Latest / Agentic Conversations (formally mlops.community) / Inferring Creativity // Nick Hasty // #198

Transcript

Inferring Creativity // Nick Hasty // #198 · Agentic Conversations (formally mlops.community)

0:00
Hi, I'm Nick Casey. I'm a VP of Product Machine
0:03
Learning at Giffy and I take my coffee with cream and monk fruit
0:09
and I usually like a very strong blintz.
0:13
Darker the darker the better. Hello Hello everyone and welcome
0:17
back to another ML OPS community podcast.
0:21
You know what it is right now we are out here talking with Nick
0:26
Hasty. He was the 1st engineer at
0:30
Giffy, a little company that did things with gifts and it
0:34
absolutely blew. Up.
0:37
We talked all about his journey into AI product from the web dev
0:43
space. Actually, before that even we
0:46
got into a little music side chat tangent, because as you
0:51
know, I love playing music and I love making music, and so if I
0:55
have a musician on here, we're probably going to do a little
0:59
bit of a tangent. So we had that happen.
1:03
We had him talk about his journey from music to web dev,
1:07
then join it in this little startup in 20.
1:11
I think you said 2013 when it was just like one other person,
1:16
maybe the founder and him and how he has been able to grow
1:23
with the company and the company has exploded as we said.
1:27
But we all know Giffy. It was like, I don't know if
1:33
it's only me because I am a millennial.
1:37
I use that all the time. I would just reply in comments
1:41
with gifts. Now I've had to start toning
1:44
back my gift usage because apparently it makes me look old.
1:49
But I love my I love me a gift. I do.
1:53
I love me a gift. And so anyway, Nick also talked
1:57
to me about the way that Giffy is using AI and how that has
2:04
changed from 2013 until now, how they were leveraging AI for
2:11
their semantic search capabilities back in the day.
2:14
And that is a cool little story. And then what they're doing in
2:17
their R&D department for AI these days.
2:22
He also filled us in on all kinds of cool stories over the
2:26
years that he's had scaling up with Get All Giffy.
2:31
I mean I feel like this is one of those that is very near and
2:35
dear to my heart because I should used it so much and love
2:39
using it. Last little bit about Nick
2:42
before we jump into it. He has been helping startups and
2:48
it is cool to see how he talked about what he advises the
2:52
startups that he helps these days.
2:54
He's mentioned that some of his former coworkers have gone on to
2:58
start things or people that have worked under him.
3:02
And it's really nice to see what he said now versus maybe even
3:07
just three years ago. All right, I think that is it.
3:12
Let's get into this conversation with Nick.
3:14
And if you liked it, leave me a gift as a reply.
3:17
That's the only way to do it. I think you gotta reach out to
3:21
me and just send me a gift and I will know that you listen to
3:24
this and you enjoyed it and give me your funniest gifts.
3:27
I want to see the best gifts that you've got.
3:29
I don't even care. No sensors involved.
3:33
Give me some weird ones. OK, enjoy this.
3:37
We'll talk to myself out of the house today and my wife, the
3:50
reason I went out of my house was because I was looking for my
3:53
phone and I thought it was in my car.
3:54
And so I went out of the house and went to look for the phone
3:58
that was in the car. But turns out my wife took the
4:01
phone and the keys that I grabbed and I left the doors
4:05
open and everything. But it's a little bit stormy
4:07
right here and boom, the the wind blew it closed.
4:12
So basically I I think ohh man my phone, it's in my car and I
4:17
walk outside, leave the doors open, grab some keys, don't
4:20
really think much about it. And the storm blew the door shut
4:27
and I realise when I get down to where the car is parked, there's
4:30
no car. And so the phone was, yes, it
4:33
was in the car, but my wife had taken the car and so then I was
4:37
like, alright, well I can't do that two factor login that I
4:41
wanted to do with my phone and they were texting me, ohh wait
4:45
to do that one. Then I get back to the house and
4:50
I realised that the doors had been shut from the storm and
4:54
then I'm like no problem whatever.
4:56
I got the keys and I grabbed the keys.
4:59
Our neighbours are going out of town and they left us their keys
5:02
and those are the ones that I grabbed.
5:05
And comedy and others. So then I was sitting there and
5:09
I went to our neighbours and I was like, can you text my wife,
5:11
tell her to come home because I'm, I'm locked out, blah blah.
5:15
She doesn't come home, she's not coming home.
5:17
And I'm realising, hey, we've got this conversation we're
5:20
supposed to be having and look up on our balcony of my.
5:26
I wonder if one of those doors is open.
5:28
So now I gotta figure out how to scale up the side of this
5:30
building. Wow.
5:32
And get on to the balcony. Managed to do that, climb onto
5:35
the balcony. The doors aren't open though,
5:37
except here in Germany. There's this like setting for
5:42
the doors where you can put it so the top folds out.
5:46
I don't know if you see. That so it's not.
5:47
Like an open. This way it opens from the top,
5:50
yeah? And.
5:50
So you crawled into that? Yeah, it goes and it opens from
5:55
the top and then there's there's another door right next to it,
5:58
but the door handle is on the other side.
6:01
So I'm trying to reach for the door handle.
6:02
Trying to open. That doesn't work.
6:05
And we've got this wardrobe outside on our balcony, and I'm
6:10
looking for tools that I can use to, like, latch on to that.
6:14
I end up finding a way to push open the door handle and boom,
6:19
it swings open And I was able, yeah.
6:22
So might as well just call me MacGyver for the rest of this
6:26
conversation. It's kind of nice to have those
6:29
tests every now and then, right? Or you're like worst case
6:31
scenario and I could figure it out.
6:33
I got it. You don't have to wait.
6:35
And I wouldn't have had any way to let you know that I've
6:39
totally blown you off, you know, So I would have felt horrible
6:42
because I wouldn't have had my phone and I wouldn't have been
6:45
able to get to my computer, which was inside.
6:47
And so I was like, I really got to get to this conversation that
6:51
was like the only thing that I had slated for today.
6:53
So anyway, man. You made it nice.
6:56
I made it and I'm glad I did. So now for those who are tuning
7:00
in, you've been working a giffy for a long while.
7:03
Giffy for those who do not know, is the millennials favourite
7:08
thing to do on Slack and or any social media.
7:12
I heard that it's not as cool anymore for the Gen Z, but I
7:16
still love it and I abuse it along with all of my other
7:19
friends. And I'd love to just dig in in
7:23
the next 30-40 minutes to your experience being the 1st
7:29
engineer at Giffy. And then yeah, going from an
7:33
individual contributor, being an engineer on the team, having to
7:38
wear many hats. Because in a startup we all know
7:41
that totally everything's on fire every day.
7:44
And then as the the organisation matured, you became your role
7:49
probably solidified more and you started leading teams and you've
7:53
been doing all of this with the initiative of using machine
7:59
learning. And so I think it's super
8:01
valuable to hear about a, how you've been using machine
8:06
learning, how that's changed over the years, be your own
8:09
personal journey and how that's changed over the years.
8:12
And then yeah, really, what gives U to these days and how
8:16
machine learning work? And so let's just start with the
8:19
idea of where did you come from, like what you joined give you 10
8:22
years ago. What were you doing before that?
8:25
Yeah, Tilly, Well, one one thing beforehand, gifts are definitely
8:28
a millennial thing, but our, our most popular content now is
8:31
stickers, which is like what people layer on top of videos
8:34
and we power stickers in a bunch of different applications that
8:37
people use. So that's kind of like so
8:39
there's the sneak, there's how we get into the the youngins of
8:43
Chintzy. But yeah, so I got kind of a a
8:47
broad background on how you grew up in North Georgia, kind of a
8:51
rural area, and there wasn't much computer stuff happening
8:57
here. Surprise, surprise.
8:58
There was a lot of, a lot of football and a lot of church on.
9:02
But I was, you know, kind of not in that zone so much.
9:05
And I got into music and was playing in like bands through
9:09
high school, like punk and even in like a ska band and stuff
9:12
like that, because that was what was popular at the time.
9:14
And then I went to College in Georgia and I was an English
9:18
major and I was in Athens, GA and I only went there because I
9:21
wanted to play music. I just, that was the only thing,
9:23
you know, that was like my my overriding belief is that music
9:26
is the best thing that anyone can do with their lives.
9:28
And so I followed that for a long time.
9:30
But then as I got a little older, you know, I was like, OK,
9:32
I gotta make money and like, maybe I want to have a family
9:35
one day. And so I I went to a grad
9:39
programme in New York at NYU, it's called an interactive
9:43
Telecommunications programme, ITP.
9:45
And what they do is take people from broad backgrounds to kind
9:49
of throw technology at them. And I had some music tech stuff,
9:52
you know, already kind of under my in my wheelhouse.
9:57
But it was an ITP where I really started doing things that really
10:01
challenged me, that I wasn't familiar with.
10:03
You know, everything from like circuit design, you know,
10:06
prototyping to programming. And, you know, I had to bash my
10:10
head against a lot of this stuff for a couple of years.
10:14
But, you know, it kind of broke through And all the assumptions
10:16
I had about myself, of me not being that kind of person, you
10:20
know, I'm like the slit guy. And you know, critical theory or
10:22
whatever, like that was like, I can't do programming.
10:25
And I was like, ohh, wait, actually I like this.
10:27
You know, I'm pretty good at it. But, you know, I kind of stuck
10:31
with like my principles a little bit.
10:33
And I was still really into art. And so after it, I started
10:36
working for this organisation called Rizzone, which is based
10:40
on the new Museum of Contemporary Art in New York.
10:42
I've been interning there and I was a fan of the place because
10:45
they were there was a nonprofit arts organisation that was
10:49
devoted to Internet art. And so it arouses the 90s.
10:53
And they had been promoting and even archiving, you know, people
10:58
doing weird things on the Internet, like weird websites,
11:01
weird e-mail stuff, you know, just people kind of like taking
11:04
apart and tinkering with what constitutes the Internet.
11:08
You know, we're we're we're websites with JavaScript like
11:10
Open infinite windows are like, you know, you look like you look
11:13
in the source code and there's like bomb schematics and stuff
11:16
like so cool stuff like that. Basically, it was NFTS before
11:20
NFTS, yeah, yeah. Yeah, yeah, true.
11:23
I mean, those ideas were discussed in that community and
11:29
before they became a popular thing, you know, and that was a
11:32
lot of fun. And in the process, I I had to
11:35
do a whole, whole lot of web programming because I was the
11:37
only engineer at this little nonprofit.
11:38
And and we had a lot of responsibility.
11:42
And also we like got really familiar with gifts because you
11:45
know, gifts are these things that are native to the Internet,
11:48
right? There's this, there's this media
11:50
format that's like been there's this day one.
11:52
And people have been doing like, you know, art through gifts,
11:55
like gifts was their medium that they made it through.
11:57
And I had to archive a lot of that stuff and I had to learn
12:00
how they worked. And so, you know, that was four
12:04
years or something. And then after that I was doing
12:08
a lot of freelance stuff, just kind of helping startups in in
12:10
various in New York City, you know, just doing early stage
12:14
stuff. And I had a friend who was with
12:17
this kind of incubator, slash Mitre Capital on an organisation
12:23
called Betaworks and they're still around, pretty big in New
12:25
York and and they were like, hey, want you come over here.
12:29
We got this new programme called Hackers and Residence where
12:32
we're pulling in, you know, kind of serial entrepreneurs and
12:35
letting them build something. They need some help.
12:37
And I was like, OK, cool. Like, what's working?
12:39
Who's who's there? And they're like, what is this
12:41
guy Build me GIF search engine. And I was like, OK, yeah, that's
12:45
fun. You know?
12:45
That sounds cool. And so I came and met and this
12:48
guys, Alex Chung. He's the founder of Giphy.
12:51
And you know, I met him that day and he kind of get it a quick
12:54
interview about how I would do stuff and some technical stuff.
12:58
And, you know, he was like, I did good, I guess.
13:00
And he was like, OK, can you start tomorrow?
13:02
And I was like, yeah, let's do it, you know.
13:05
And so, yeah, so that was, you know, this is like February of
13:09
2013. I think he built the first Giffy
13:12
prototype in November of 2012 and it was just like a website
13:16
doing a you know full text search against MySQL database,
13:21
you know with a bunch of crawl Tumblr data And but it was like
13:24
super popular and people were crashing it, you know because
13:26
they were because they were they were using it.
13:30
There was like this demand, right?
13:31
There was very clearly a demand. And so we started building this
13:35
thing. And you know, he was coding and
13:37
I was coding and I was doing the website, all of that stuff.
13:39
And he was doing all of the API. And yeah, you know, it was just
13:42
one of those things. It was like if you keep building
13:45
it, they keep coming and and you know I think what we, you know,
13:51
our operate. You know what we were kind of
13:52
like working on was the idea that there's something happening
13:56
in gifts that is very unique and distinct and kind of new and
14:00
that is that people are using them to communicate.
14:03
And we we saw this in action on, you know, a lot of different
14:08
message boards and websites like that.
14:10
The rural popular back in the daylight, like Tumblr for
14:12
instance, you know, her people were post something and then
14:15
people would put comments in the gifts, you know, put gifts in
14:18
the comments, you know, as like a reaction or a response.
14:21
Or sometimes people would post something that was a gift and
14:23
then they would have some caption underneath that is like,
14:27
you know, something make it there, making an analogy between
14:30
what's happening in the image and then what's in the text,
14:32
right? It was like, usually real funny,
14:34
like real silly stuff. It really this is cool.
14:36
This is a really interesting emergent behaviour that's
14:39
happening here and like we should just get this everywhere,
14:42
right? Like this needs to be in every
14:45
place that people are on the Internet or any place that
14:47
people communicate. We want to have them give them
14:50
the ability to communicate with in this format.
14:54
So that's what we did. And you know, we plugins for
14:58
every WordPress, you know, like every everything that people
15:02
posted. We build a plugin for it.
15:04
You know, Slack wasn't even around yet, but there was like
15:07
we chat and things like that and we would build integrations
15:09
there and we just built this network and we were pulling in
15:12
content from all over the Internet.
15:13
And the goal was to just create this big network, right.
15:15
We didn't want to create a social network, which was, I
15:18
mean, at this point in time it was still like an idea like,
15:21
hey, we're going to create a social network, right?
15:23
And people like, oh, cool, you know, but we're not going to do
15:26
that. We're going to create a network
15:28
that is a content distribution network for communication that
15:31
plugs into these different networks.
15:33
And it was funny too, because we would go higher and we'd look
15:37
for people and most people would kind of laugh at us.
15:38
They were just like give search engine.
15:40
That's stupid, you know, means whatever.
15:41
That's like, you know, they would laugh And I was like, OK,
15:44
that's fine, you know that you probably don't want to work here
15:47
then. Meanwhile, yeah, you're kidding.
15:50
Like gigantic scale and. Yeah, ohh yeah, totally.
15:55
I mean, it was like, it was, it was wild.
15:58
You know, sometimes I was like, I can't believe this is
16:00
happening to me right now because it was like one of
16:03
those, like, hockey stick, parabolic growth things, right
16:05
where it was just like you. It's like here's our, you know,
16:08
here's how many gifts we're serving and it's just like Boo
16:11
Boo, Boo, Boo, but, you know, straight up and we're just
16:13
constantly deploying our infrastructure.
16:15
I mean I was kind of you know, full stack person.
16:18
I I was like really struggling to keep U and to learn and to
16:22
like how do I provision all this stuff and write scripts.
16:25
I mean, virtual machines were still kind of new, AWS is still
16:28
kind of new, and I was. Like brunettes was not a thing
16:30
at all. Yeah, like Auto Scaling wasn't
16:34
totally there yet and it was like, it was like it was a wild
16:38
ride. But yeah, you know, it's a lot
16:41
of fun. I like those kind of situations.
16:43
And you know, like what you said earlier, you know startups or
16:47
you know, they're kind of a free for, all right?
16:49
Like you gotta have alignment in terms of it's like, I mean to me
16:54
they're like cults, you know, it's like a cult, right?
16:56
You have a cult leader, kind of the founder.
16:58
And then everyone has a really, really believe in that founder
17:00
and believe in the idea. And that gives you a certain an
17:04
irrational principle or an irrational set of beliefs that
17:07
you adhere to that guide your actions and guide your
17:10
behaviour, right? Because all things being said,
17:13
there's no reason why any of that should work.
17:15
You know, like left alone, any of these things would not emerge
17:19
and by itself, right? So it's like, OK, we got to make
17:22
it happen and down. So yeah, you got to do random
17:25
stuff. Like, yeah, you may be
17:26
programming, but then you may have to like, you know, run a
17:30
Reddit community on the side or letting you know or like, you
17:33
know, talk, do BD development or like whatever.
17:35
Because if you don't, it's not going to happen, yeah.
17:39
Ohh, that's such a good so many parallels there.
17:42
I mean I first want to just point out I feel like I got into
17:46
tech the same way as you. I you know, music lover at heart
17:49
and wanting to just play music or the rest of my life.
17:53
And then my daughter was born and I'm like oh I got to go make
17:56
money. Can't be doing this music thing
17:58
like 24/7 anymore. So had to take a long hard cold
18:05
look in the mirror and get into tech.
18:08
And so there's that piece. But then I really like how you
18:12
were able to adapt because I feel like a lot of people aren't
18:16
able to hit those scaling issues or hit those issues where it's
18:22
like ohh. Now I need to go figure out like
18:25
how to do this. I've only done web development
18:27
but looks like I need to be more full stack.
18:30
Let me go figure that out. Where do you think you went?
18:34
Right where others may fall off. You mean me individually or like
18:39
the company? You individually.
18:41
I think I kind of tend to operate with a certain like of
18:47
recognition of all the uncertainty that I have.
18:51
And that uncertainty is like really strong.
18:54
And so I'm just like, I don't understand this.
18:56
I don't know enough and it takes a lot for me to be satisfied to
19:00
actually know something. Umm.
19:03
And so you know during this time period, it's just this kind of
19:06
constant, unrelenting search for pupil for information online for
19:11
whatever it takes to figure out how to solve this problem and
19:15
never assuming that, you know, it's not just going to come
19:19
immediately, right? Like I'm not going to I'm I'm
19:21
not just gonna magically get struck by lightning and figure
19:24
this out, right. Like I gotta get out there and
19:27
hustle and and and you know do what it takes to solve that
19:30
patch that problem is send the time and effort to research.
19:34
I mean, you know and then and then try to apply and then you
19:37
know a lot of times it doesn't work because you have imperfect
19:40
information and you're doing this.
19:42
You're like oh shit that didn't work out.
19:43
You quick thank. You.
19:44
How do you move? What else?
19:45
What else? You know?
19:46
What am I? What do I don't?
19:47
What do I not get here? And you know, having the
19:50
pressure of angry people on Twitter saying, well, gifties
19:55
down or something, you know, like that she's like, huh, I was
19:59
like, yeah, you're sweating. Mom comes after you, Yeah.
20:02
That's so totally, yeah. The anxiety that must induce the
20:06
hundreds of thousands of people you're letting down, like they
20:09
can't react in the way that is most natural to them.
20:13
Yeah, Tilly, yeah, we, you know, we took it with a grain of salt.
20:16
I mean, we were always, like, acknowledging the absurdity of
20:19
the situation. You know, like we're working.
20:21
We're getting paid to like give people gifts.
20:24
And if the gifts don't flow for a little while, like, you know,
20:28
all the orphans and the sick people in the world, you know
20:31
they're still, they're not, you know, we're we're not we're not
20:33
killing anyone or not her, you know.
20:36
So we're lucky to have a, you know, the stakes were high, but
20:40
they were different than a lot of other people who were doing
20:42
things, you know, so. Yeah, that's so good.
20:45
And now one thing that I'm interested in too is the
20:49
trajectory of AI. And maybe we can start by diving
20:53
into like, how did you shift from being an engineer to
20:58
growing more into product and AI?
21:01
Yeah, totally. So I think I'm, you know, I
21:06
would consider myself more of a product engineer in a way, which
21:10
is great when you're a startup early on and you need someone
21:12
who's like can kind of imagine the product and then build the
21:16
product. And it may not be the best and
21:18
it may not be the most efficient, but like it, you
21:20
know, as long as it works and it doesn't, you know, kill the user
21:25
experience and off the need to ship it.
21:28
But I think it's giffy. Started scaling.
21:29
You know, we started to have to we hired like real engineering
21:32
problems that we had to solve. You know, like ridden like like
21:35
as soon as we hit these integrations with major, major
21:38
social networks and you know, Facebook was hit us up and
21:42
they'd be like, yo, you got to shave 30 milliseconds off your
21:44
response time. You know, like, you know, that's
21:48
like, OK yeah, yeah. It's like, OK yeah.
21:51
So then you know we those problems started happening and I
21:55
addressed those the best I could have been you start hiring to
21:57
serve those things. But I wouldn't say that's like
22:01
my zone right? Like I can I can try to figure
22:03
it out but I'm not gonna be like this isn't my hay place Rs.
22:07
E&O. Luckily, I started just moving
22:10
into like what we called special products in R&D at first.
22:14
You know, once we got the other engineers in there to help with
22:16
some of those scaling things and focusing on those things and I
22:19
was like, I don't really want to do this, you know, I want to do
22:21
something else. And so you know I had a pretty
22:24
good sense of the problems of the business in general on
22:28
started trying to figure out what are things we could build
22:30
internally that will make things easier for the company to flow
22:33
and the company to be more efficient or what are things we
22:36
can what are experiments we can do to you know showing new ways
22:40
you can interact with the product and.
22:41
I'll ask stickers. Yeah, yeah, yeah, yeah.
22:45
So one of those things was like, you know, search at this time
22:48
period was super powered by tags.
22:51
You know, everything was tags, tags, tags.
22:53
And we would pull in content from the Internet.
22:56
And then we had like, you know, a team of like 5 or 6 editors.
22:59
They were like curating and tagging stuff, right?
23:03
But the more content we pulled in, because I had written a lot
23:07
of web crawlers and stuff and I was just kind of vacuuming in as
23:09
much as we could. Clearly, you know, that problem
23:12
that that doesn't scale. And so I was like, OK, well
23:15
there's there's this new thing, computer vision mill, you know,
23:18
there are people who are, you know, finding labels for stuff.
23:21
And I was like, this is, this is what we got to do, you know?
23:25
So I started looking into that, researching that.
23:27
I didn't really have the technical background there.
23:30
So I sort of you know looking for 3rd party solutions or like
23:34
other ways of just spinning stuff up, you know knowing that
23:37
there was number way wheat the company continues to grow
23:40
without having this stop. So how do we get it in and start
23:42
moving it and you need leveraging it, you know and back
23:46
then it wasn't so great. I mean this is very general and
23:49
like object detection stuff, but so it wasn't you know in our in
23:54
the metadata we needed tended to be hyper specific and very
23:57
culturally relevant and abstract and that kind of stuff.
24:01
But we stuck with it and then there's like other models
24:05
started coming out that did more things, we were able to lose
24:08
leverage does like for instance, where did that?
24:10
You know that that came out and that was primarily for NLP but
24:14
and we were able to leverage it for content recommendations and
24:18
we've and it's still like used it's still outperforms like
24:21
graph networks and other things we've tried out.
24:23
You know we had the idea or read the idea and adapted it to like
24:27
take user information like user session data anonymized of
24:30
course. But like hey they looked to get
24:31
us GIF. They did this search and then
24:34
they clicked on that and put that at treat that like a
24:37
sentence in a corpus of text and then feed that into words.
24:41
Avec and it learns semantic relationships between like we
24:46
would use GIF IDs, not gifts, but the idea of the GIF and
24:49
learn the semantic relationship between like the GIF and the and
24:52
the query and the click and these kinds of things.
24:54
And you build like a really great embedding of like pop
24:58
culture and communication and expression and emotions along
25:04
with images. You could even do, like the kind
25:07
of arithmetic stuff. Or you could take like, you
25:10
know, like what? Like king is king minus man plus
25:13
woman equals queen. And for us it was like Aragorn
25:16
from Lord of the Rings minus somebody plus a woman.
25:19
And then it gets like Beyoncé, like Queen, did you know, like
25:23
it was like a very like, like it totally worked like that was
25:26
like you could literally, you know, traverse the vectors like
25:28
that. Wow.
25:29
It was super fun and. This was pretty vector database
25:33
explosion. Yeah, right now, right?
25:35
Exactly right. This is like 2018 or so.
25:39
And so we were still, you know we're using like K&N lookups for
25:43
these kinds of things and pressing them in that O.
25:47
So that was a lot of fun. And I think it when I started
25:50
seeing those things I was like wow this is super interesting
25:54
creative of these models can learn these things like on you
25:59
know maybe in the world we live in now all these problems seem
26:04
so solved. But NLP was still struggling
26:06
with like you know these kind of things back in the day
26:09
especially like language is messy and slippery and you know
26:13
different things have different meanings and and it all depends
26:17
on the context right languages and ultimately is just a
26:20
construct of how people communicate and how our brains
26:24
learn the context of what we communicate right.
26:27
And that's kind of the the foundation between a lot of
26:29
these models. And I was like like word of it
26:32
was all about, you know, context.
26:34
What's the context in which the thing appears in that?
26:36
That's still very relevant, obviously to like things like
26:38
attention and these kinds of things, but depending.
26:42
So the context is learned from the data, right?
26:45
And if your data is emotions and memes and moods and feelings and
26:49
images, then yeah, it's going to learn stuff.
26:51
And it's going to learn the difference between, like if you
26:54
have a search shown like good morning, right?
26:56
Like good morning. Sounds like simple, but there's
27:00
like 10 different ways to say good morning.
27:01
There's like, ironic good morning.
27:03
There's like, you know, sort of sarcastic.
27:05
Good morning, Good morning. Right.
27:06
Yeah, exactly right. Yeah.
27:08
Totally. Yeah, exactly.
27:10
So many different ways. So many different ways, right?
27:12
So when someone searches Good Morning like you don't know what
27:16
exactly version of Good Morning they have, right?
27:18
What's their intent? But these networks kind of
27:20
helped us learn these things. And then we could present a wide
27:23
variety of interpretations of this really simple concept.
27:28
Yeah. And then when they.
27:29
Click in you see the ohh this is also.
27:31
Here's a few that are similar to this.
27:33
Yeah, true. Yeah.
27:34
And it's all reinforcing, you know, the more they click, the
27:37
more these models. But yeah, I think I was like,
27:40
wow, these, these kind of, you know, works very simple
27:43
relatively speaking, compared to a lot of things.
27:45
But it's learned these super complex relationships like irony
27:50
and sarcasm. You know these things
27:52
linguistics used to say. Ohh, you know, there's no way
27:54
you can encode that into programming and stuff, right?
27:58
Because there's just too much complexity or too much slippage.
28:00
But if everything's just statistical, then then boom, you
28:04
can solve it, you know? And so I think that got me
28:07
really inspired and really excited to see that happen.
28:10
What do you think is one of the reasons that you would call
28:13
yourself or you would identify as a product engineer more than
28:18
anything, Yeah. I mean, I like to, I like to
28:21
make things for people to use, you know, like I like, like if I
28:28
if I have an idea for something, I want to see it emerge.
28:31
And my ideas for things are usually things that people
28:35
interact with. You know, like in music too.
28:37
You know when you play music, you play it partially for your
28:39
own, like, joy. But it's also you're sharing an
28:42
experience, right, with other people.
28:45
You're creating something inside of you with some with other
28:49
people, and they think that, you know, you can think of a product
28:51
like that too. It's like there's this thing
28:53
that could exist in the world that could make people feel good
28:55
or feel joy or feel hay or maybe better their lives in some way.
28:59
Yeah. So you say you you want to see
29:03
it happen. You know there's an itch, right?
29:05
There's like, there's a pain in the world, or there's a problem
29:07
and I have it itch and you think you can solve it, you know?
29:09
Yeah. Do you feel like that has
29:11
changed a tonne? No.
29:13
Since the advent of MLM's and this whole boom?
29:16
Because I feel like just from my perspective, not to bias your
29:20
answer or anything, but what I've seen is that the product
29:25
engineers who understand a bit about how API's work, they can
29:29
hack together something pretty quick, are able to stand U MVPS
29:34
that they don't have to now wait for the mill team or the
29:38
platform team or the data team to actually get going.
29:41
They can just say hey look here, if we did this I think it would
29:45
be useful. Here's a quick demo of it.
29:48
Is everyone else on board, You know, SO like how in your
29:53
experience has it changed because of that, if any?
29:58
Yeah. I mean, I think you're
29:59
absolutely right. I mean, shipping ML products has
30:03
historically been difficult, kind of tedious, right.
30:06
It's like you got to gather the data.
30:08
Do we have the right data? You spend a tonne of time like
30:11
making sure that the model is going to build a representation
30:14
that is aligned with what you want, yeah.
30:17
Hey everyone, my name is Aparna, founder of Arizona, and the best
30:21
way to stay up to date with ML OPS is by subscribing to this
30:24
podcast. And that's the whole thing into
30:28
itself. And then you got to find the
30:29
right network. And then you train and you
30:31
evaluate, you know, And then it's like, OK, what's the
30:34
interface? Like, how does this thing live?
30:36
And then before that, like, how do you deploy, you know, how do
30:39
you prop this thing up? Because it's like we've got this
30:41
huge binary file, Like that slows hell.
30:45
Like, how do we. So yeah, it's a whole world,
30:47
right? You don't even know at the end
30:49
of the day after doing all of this if the end user is going to
30:52
find it valuable. And you got to go through all
30:54
that. Crap, just to figure out that
30:56
ohh yeah, this feature isn't even that useful.
30:59
Totally. I mean, exactly.
31:00
So, yeah, I mean, you know, these things alms, you know,
31:04
they're, they're so powerful and they're so easy to, to do things
31:08
that were historically very complicated even.
31:11
Like, you know, like you can, you can use a sophisticated LLM
31:16
to do really interesting and novel analysis on data and like
31:21
text in real time, right. And like, it's not just about
31:24
producing text. It's like, hey, here's something
31:26
the user said, you know, evaluate this and classify it,
31:30
right. So if you have the right prompt,
31:31
you can build a cool classifier, right.
31:34
That takes in some really difficult text or not even
31:38
difficult. Like, you know, it's one thing
31:41
where if you're building stuff for people and they have to
31:43
press a button and that button is true or false, right?
31:46
If you're working with speech or language, then there's a whole
31:49
lot of different ways to say true or false, you know, like
31:55
yes, maybe our sure thing or whatever, you know.
31:58
And then you can go the true false classifier based on
32:01
language you know, using an LLM. And that's, you know,
32:04
historically, you might have some kind of crazy substring
32:07
list or whatever, right? You're checking against, right?
32:09
But it's never complete, right? It's never, it's never enough.
32:13
But these things, you know you can.
32:14
So you can, you can, you can put all these things together and
32:17
you can build really amazing. Yeah, exactly what you're saying
32:20
is of stuff, right? Because I think what's
32:23
interesting is that the the interfaces are moving away from
32:26
being purely boolean in a way, buttons and stuff, to being more
32:30
statistical in nature, which is more like reality itself.
32:36
O We can build interfaces that are more natural to to people, I
32:40
think you know. And I I also wonder, when you're
32:44
thinking about building products now, how do you evaluate?
32:49
Should we use an LLM for this use case or should we use this
32:54
older or traditional or smaller model or boring less sexy thing
32:59
because we've got years of use out of it?
33:04
It's battle tested. If it ain't broke, don't fix it
33:06
type thing. Yeah.
33:08
I mean to me it's all about what are you building, right?
33:11
And what are the what is the value you're delivering in the
33:13
product and if and and in that sense the tools are chosen out
33:20
of necessity and not because you want to do some use some tech
33:24
right. If you got something battle
33:26
tested that works in fulfils the need, yeah, use that.
33:30
You don't need an LM, but if you have something that there's no
33:33
existing tool for or using an LLM will save you.
33:39
You could do it the old fashioned way, but you know, you
33:42
would have the pain that we kind of just described in that
33:44
process then yeah, use it. I mean the only with alms is
33:47
that you know, they're kind of slow.
33:49
You know, there's a latency issue and there's the
33:53
scalability issue. I mean, I think all those things
33:56
will go away overtime. That's kind of the nature of of
33:59
tech. But I think if you're building
34:02
something that's purely LLM based right now, then you're
34:04
going to have to contend with those things.
34:06
Yeah, you're getting that call from Facebook.
34:08
And there's been. Shaved 30 milliseconds off and
34:11
you're like I can't do that. The opening eye response time is
34:18
anywhere from three minutes through 30.
34:20
Exactly. So.
34:21
So, yeah, that's what I mean at the thing you're building, if
34:25
you can build in to the end product, if you can build in the
34:30
expectation for a delay, right, You know, yeah.
34:34
And the experience? Then maybe it doesn't matter so
34:36
much. So again, you know, it's all
34:41
about what are you building and what are you trying to deliver.
34:43
I think yeah and but, but, but now there's so much cool stuff
34:48
you can deliver, things that could never have been done or
34:51
have, you know, much more complexity or interesting in
34:55
states than you ever could. You know, it's a great time, in
34:59
my opinion, to be an engineer. Yeah, yeah, it's amazing, you
35:02
know. It's incredible and especially
35:05
someone who thinks about product and yeah product focus product
35:09
minded. And I I also wanted to ask you
35:13
thinking about building products and especially in this AI with
35:19
space that you play in. What's your opinion on like the
35:27
intuition of, hey, I know this product is going to work well
35:30
versus we need data on if this product is going to work well or
35:36
we need to see more examples of this working well before we
35:41
actually move forward with it. And especially because, yeah,
35:44
you're you're building at a gigantic scale, right.
35:48
So yeah, it's something that I imagine you can, if you want,
35:52
you can AB test like crazy or you can do a lot of experiments.
35:56
And so you've almost got like all that at your fingertips.
35:59
How do you look at leveraging that when building these AI?
36:03
Yeah, yeah, totally. I mean, I mean I think kind of
36:07
have to have both. I think you know when you if you
36:11
come U with something new, you know it can come from data.
36:16
You know, like you're like there's some like you're very
36:20
analysing a bunch of user data. Like if you have a pre-existing
36:23
product and there's a bunch of user data about how people use
36:25
that product. And you can examine that data
36:28
and then say, OK, there's a pain point here or or there's some
36:31
kind of like narrative in here that's unique and different.
36:34
And you can isolate and pull that out and then build around
36:39
that narrative. You know, that's like A1
36:41
data-driven away from starting from scratch.
36:44
Or it's like, you know, intuition is just in my opinion,
36:50
you know, like you're this inference based on your brains
36:55
collected wisdom and knowledge, right?
36:57
So it's like O if you know a space really well or like a
37:00
domain really well, you can say, OK, I have some intuition about
37:04
what could be a good product in this space and then you start
37:07
building and then you have to test right?
37:08
And then I mean what you build and what your intuition is
37:12
though oftentimes is very different.
37:15
You know what the end result is versus what you think it needs
37:19
to be out there oftentimes is very different and sometimes
37:22
you'll be right and you'll be like I was exactly right.
37:24
That's a home run like this. What I had in my head is he's
37:26
like what people wanted. But like that's very rarely the
37:30
case and when you are thinking that these scales, you know
37:34
you're you're thinking about distribution as abusers and
37:37
you've got to find what's the, what's the product that makes
37:42
the top of the bell curve A is to get the most people hay per
37:44
se or unless you want to. But sometimes you do have to
37:47
engineer for the for the outliers in these kinds of
37:49
things. So I think some mixture of both,
37:51
you know and this is kind of well worn territory with product
37:55
development and startups, you know, it's like user testing,
37:57
continuous you know, testing and product development and that
38:01
kind of thing. I think that's all true.
38:03
I mean, I think you can. You can't dilute yourself
38:06
though, right? And like, I think you have to
38:08
have your North Star. Like what am I actually trying
38:10
to build and what what's the experience I'm going for here?
38:14
And if you end U doing too much testing, I think you could say
38:17
you could end U with some kind of, you know, a product that's
38:22
delivered by, you know, too much collective information.
38:26
And it's kind of like, you know, doesn't make anyone upset but
38:30
doesn't really make anyone happy, right.
38:32
Yeah. And that could be a bad thing.
38:34
So I think you got to have both. But you got to stick to your
38:37
guns. Really.
38:37
You got to stick your vision and your vision may change.
38:41
So it's like what you build may not be the thing, but as long as
38:44
whatever you end U with is delivering the same kind of
38:47
value or the same kind of joy to the user that you wanted to,
38:50
then, then you win. Yeah.
38:52
I like what you said right there.
38:54
This product development that ends U not pissing anybody off
38:59
but not making anybody a super fan.
39:01
Yeah, and how it's like you playing it safe isn't getting
39:06
you any accolades. In a way, yeah, but you're also
39:11
not going to get fired. So hopefully I Yeah, it's a
39:17
funny one because it's like you almost have to fight against
39:20
that inertia to just do what you think is going to work and
39:26
because or what you know will work as opposed to taking a
39:29
bigger swing and going for something else.
39:32
And especially with, I mean that's the idea with that you
39:38
were talking about with the LMS and being able to do so many new
39:41
things. Yeah.
39:43
And being able to incorporate AI into your product in all these
39:48
different ways. Yeah.
39:50
This is a great segue though into talking about what you all
39:53
are doing now at Giffy as opposed to you told us about
39:58
back in the day. This word to vec is incredible.
40:00
I love that. But.
40:02
How has it evolved? If it has, maybe you.
40:06
Yeah, for sure, says MLB team that I've been running for a
40:10
couple of years. You know, it's always pretty
40:11
small and scrappy, but we've been able to do a lot of
40:13
different things. Back in 2018, we had a network
40:21
that we released, Snooping Source was like a celebrity
40:23
detector. And it's like you just said,
40:26
convolutional neural network that we trained on a bunch of
40:28
celebrity information and you could upload a photo and it
40:31
would detect, you know, what celebrity you looked the most
40:33
like, what celebrity it was. And you know, that was like a
40:36
fun, giffy S kind of product thing.
40:39
But in terms of, you know, inside the company, like it's
40:42
used all over the place. You know, we use machine
40:43
learning, these kinds of things. Obviously we do search, right?
40:46
Like so any company, anyone doing search now is and not
40:51
using Mill and the search is going to miss out.
40:53
So you know, we used that same model I had mentioned earlier.
40:57
We use embeddings from that in the search.
40:59
We do, you know, just various kind of vector embeddings and
41:03
those kinds of things to help help us find better matches, you
41:07
know, beyond like the whole goal is like, you got this sequence
41:10
of images moving, you know, what are all the ways you can extract
41:14
metadata from that? So you got to be scrappy and
41:18
come up with creative ideas. And then we also have things for
41:21
like tagging, like when people upload content.
41:25
And we have a model that we've created that takes in the image
41:28
of the chief and then recommends tags based on GIF's, tagging
41:32
structure and popular queries that allow people to tag that
41:35
content. So it'll help show up in our
41:37
search on. But yeah, you know ever since I
41:43
mean so like if he in and of itself is like at least the way
41:46
most people use it which is in Slack or some sort of
41:49
integrations which sort of third party integration like you put
41:52
in some text and you get an image, right.
41:55
So obviously when Dolly came out the first one we were like oh, I
41:59
was like oh that's it, you know that's the future, you know.
42:01
But you know we didn't have all that resources.
42:04
So but we've been kind of steadily playing with that the
42:07
best we could and we've been focusing on generative stuff a
42:10
lot. We don't have any products that
42:13
are like release candidates right now, but I mean, we do
42:16
have it internal. We built a fun internal
42:19
generative GIF maker that works just like the Slack bot.
42:22
You know, you can type in some text and then we'll generate
42:24
gifts, you know, and largely because, you know, I think this
42:29
is true any generative video product you use right now, like
42:32
everyone knows, the quality can vary, right?
42:34
Yeah. And so like, I mean, there are
42:36
some amazing ones as amazing engineering behind them, but I
42:38
wouldn't consider generative video solved right and totally.
42:42
So hit or miss too and just so hitting on different basically
42:47
how the how the model feels, I don't.
42:50
Know how to actually, yeah. I want to say on.
42:53
I wouldn't say there's something that we're like, yeah, this is
42:55
at lunch, I mean maybe if we had different resources, but but
43:01
we've been, you know luckily the open source community is great.
43:04
There's like all these different models that come out that we can
43:06
experiment with especially like animated GIF, the latest one,
43:09
you know, we've been using that for a lot of stuff in, in house
43:12
just to experiment with the anticipation that you know
43:17
people will solve this problem in the community or you know,
43:19
whatever. And we can, you know we've
43:21
always been kind of an applied machine learning like so So what
43:26
are we building right. We're building something that
43:27
could generate a sticker or generate a GIF.
43:29
Well, what's our user expectation?
43:31
Or their, you know, maybe maybe they can deal with some some
43:36
noise or some weirdness. Or maybe they can't.
43:39
I mean, we do have past experience that tells us that
43:42
people do like weird stuff, especially in Slack.
43:46
You know the early days. Of just thinking like you.
43:50
Yeah, if you know all the gifts that I've posted you probably.
43:55
I can't. I'm probably like, again, middle
43:58
of the road, but I've posted some weird shit with this.
44:03
Imagine the other weird shit you've seen.
44:06
Ohh, yeah, we had special and special collections.
44:09
So we still have just our highlights.
44:13
But yeah, like with Slack, right, early on it was kind of
44:16
very random, right? Like our collection was limited
44:19
and people would use the integration and we'd return
44:22
something and you know, it was still very random and it was
44:26
kind of random. People either loved it or they
44:28
hated it, right? Like they were like you guys
44:31
can't do search. This is highly relevant.
44:33
What's going on? Or people be like, ohh, this is
44:35
so weird. I get the juxtaposition between
44:37
what I typed in and what this image is, you know?
44:39
And then like, the whole office laughs, right?
44:42
Because some funny thing comes back, right?
44:44
And then the whole office is everyone's dealing with this
44:46
juxtaposition between this text and this image and the
44:49
conversation, and it becomes very playful, right?
44:51
So we do know that there's a threshold, you know, that people
44:56
are willing to accept the weird or the funky or the the low
45:01
quality. I mean, gifts are super low
45:02
quality historically. You know, we we largely serve
45:04
videos at this point. You know, we've turned
45:07
everything to videos. But you know, pixelation is like
45:10
fundamental to the to the media itself, right.
45:13
So like people, people's tolerances are pretty high, but
45:17
yet to your point, you know, I think once we get a little
45:19
further along with the quality and the speed and all that kind
45:22
of stuff, then yeah, we'll, we'll definitely be there.
45:25
So let's jump for a SEC. Because you also do some really
45:32
cool stuff when it comes to helping out other startups.
45:35
I mean, you are like, I guess the poster child for being able
45:40
to just get your hands dirty in a startup, as we've heard over
45:44
the last 40 minutes. Yeah, you took to it like a fish
45:49
in water. And I would love to know, like,
45:53
what are some things now that because you're working with a
45:56
venture capitalist or you're doing stuff, you're helping out?
45:59
What is it? The what's going on?
46:02
Give me the story there. And then also just like what
46:05
have you been seeing and helping different startups with?
46:10
Yeah, yeah, totally. I mean, yeah, I mean, I love the
46:14
startup, the idea, and that, you know, just being involved.
46:17
And I love helping people who have something kind of realise
46:21
it. So let them a lot of different,
46:24
you know, consulting on the side, you know, fun projects.
46:27
Some when you're starting something new and you've never
46:30
done it before, it's very scary and challenging, right?
46:33
And you're like leaping into the void because there's just like,
46:36
you know, just sting, like who knows what's going to happen.
46:39
And I got to do everything right.
46:40
So I've had people who've worked with me or worked under me.
46:43
They do stuff and then I'll just kind of, you know, help along
46:46
with their prototype, you know, just like do like, hey, let's
46:50
walk through this product, right.
46:51
Tell me what's going on. What are you thinking?
46:52
Anything from that to like actual coding or running teams?
46:56
You know, like when you when you have a startup and you're if
46:59
you're not a technical founder or whatever, even if you are a
47:02
technical founder, running engineering teams is hard,
47:04
right? Like and so on.
47:06
I've helped other people, like kind of prop up their teams and
47:08
get that organised. To me it's all about alignment
47:13
and respect. I would say like, you know, like
47:17
again, I have this like communication literature
47:19
background. Some can be pretty articulate
47:21
and like I can also have ideals if you can romantic about stuff,
47:25
you know? And I think you kind of need
47:27
that to to lead teams, you know, because like, you know, if
47:30
you're just saying, if you're just walking through a board and
47:33
saying, OK, you do that, you do this, like that's demeaning.
47:36
Most engineering is boring. Most engineers I know are really
47:39
smart, creative people, you know, So you kind of have to get
47:41
them. Kind of goes back to that cult
47:43
thing, right? Like you say, like, hey, this is
47:45
what we're doing. This is why it's important to
47:47
the company. This is where you fit in.
47:50
And This is why we're asking to do this work, you know, And then
47:54
you trust him to go solve the problem and you give engineers
47:57
the space to solve it, you know, and that gives them agency and
48:00
autonomy. And then they'll do their best
48:03
work usually. Right.
48:04
And then you've kind of find that on.
48:06
I mean, it's like collaborating with musicians, right?
48:08
Like you may have written a song.
48:10
You're like, here's a song, but I want you to kind of figure out
48:12
this part. You haven't solved that and you
48:14
let them do it and then they contributed back into it.
48:17
Right. So it's not just your song.
48:19
It's like everybody's song. It's the same thing, I think,
48:21
with with project work. And I know that sounds fairly
48:23
romantic and some some work is more interesting than other
48:27
work, no doubt about it, you know.
48:30
And I, and I personally tend to gear towards things that I find
48:35
very interesting. I've done some like SAS and kind
48:37
of fintech consultations and to not to hang out there too long
48:42
because I just could. You know, that's not my jam.
48:44
But I mean I'll, I'll but you know starting from from scratch
48:47
to be very helpful and yeah, you know to the Venture point, you
48:53
know it's just a very different landscape than it was back in
48:56
the day. And I'm in kind of following
48:57
that and talking to people in the industry and you know when
49:01
Gifty was raising and lots of people were raising on the App
49:04
Store was there, but there was still this feeling like you can
49:09
create something on the Internet.
49:10
And as long as there's demand, like just let that thing go,
49:14
right, don't worry about revenue because like you know, revenue
49:18
is like all right, revenues. Friction, you know, making
49:22
people pay is friction. We're trying to see how big this
49:24
thing can get. You know, now it's kind of
49:26
investors would give money just based on growth and I think
49:31
that's not the same now obviously since maybe that kind
49:33
of peaked in 2021. But you know, once, I mean you
49:38
know, economics is like the Fed policy changing and everything
49:41
and a lot of these bubbles, crypto and metaverse they've
49:44
they've burst and I think a lot of MCS are kind of whoops, you
49:48
know, a lot of DCS and you know, so there's been kind of a
49:51
calling I guess you could say. But you know people care about
49:54
revenue now and they care about traction with users and numbers
49:58
and I think it's harder to have an idea or whatever.
50:01
So it's like I guess I'm trying to say is it's kind of
50:03
interesting because as much as you can go out and build
50:05
incredible things now that are very much more difficult and
50:12
much more have capabilities you could have never had in the
50:14
past, it seems harder to get that VC backing right And then
50:20
things aren't as Greenfield either.
50:22
I like, you know, big tech owns a lot and it's hard to, you know
50:27
where your Apple you have a store and you take a big cut and
50:29
like you know they make the most money.
50:31
Lots of companies started, but Apple made the most money from
50:34
the App Store, right? Yeah I feel like I was kind of
50:37
going that same way. Whereas someone like Opening Eye
50:41
who have true disrespect for but they're probably trying to
50:43
follow that a store model in a way where they're like we're
50:46
going to build all the capabilities you need and then
50:49
you you're going to create apps or products on top of these
50:52
capabilities, right. Which which can be cool and
50:56
maybe you'll come U with stuff but but you know if you're in
50:59
Mel infrastructure company right now then you're probably
51:02
sweating. You know if you or if you're an
51:05
mill specific tool that does something this is kind of a
51:09
generalised tool then you're probably sweating right now
51:11
because big tech has so much money and resources and they can
51:14
eat you alive if they choose. So you know, if to build right
51:18
now, I think you've got to be really specific about your user
51:22
and you probably got to be building at that and those edge
51:25
cases a little bit, right. We're like, this isn't like if
51:28
he survived because no, no big tech companies wanted to build a
51:32
gift search engine. It's not like we add some, like,
51:34
unique technology. Yeah, exactly.
51:37
Right. It's like, well, they already
51:39
did it, right? And it's pretty good.
51:40
So let's just incorporate it into this thing because we're
51:43
not going to spin up a team to do that.
51:45
Yeah, I think that's kind of you have to look in those spots,
51:48
right, where like what does big tech not going to be dealing
51:51
with or not paying attention to and then you can kind of build
51:54
around that and we'll see. I I don't know if the idea of
51:57
these like billion dollar companies everywhere is the
52:00
same. Like you may have a lot of $100
52:02
million companies that are kind of built on top of these
52:05
different infrastructures on. So we'll see, you know, just a
52:09
whole it's a different world in my opinion.
52:12
But it's at the same time, it's very mean.
52:15
Yeah, I build, I build stuff on my own all the time because it's
52:17
just like how can you not be excited about this.
52:20
And yeah, they may not be the biggest thing ever, but like, I
52:23
don't know, you could probably do well and make some money and
52:25
maybe get some other people to help you out and they make some
52:27
money and then maybe do it again, you know?
52:31
Yeah. Especially these days, being
52:34
able to leverage what Open AI has put out.
52:37
Exactly. It's incredible.
52:40
It is a little bit like, I know I shouldn't be using it because
52:44
I'm just going to get myself a lock in and I'm going to create
52:47
a better ecosystem for them. But they make it so damn easy.
52:50
Yeah, there's that paradox in a way.
52:53
And and then the other thing that you said, it's is quite
52:57
funny because I imagine back when Giffy was raising money, a
53:02
seed round was probably like 1,000,000 bucks maybe or
53:07
500,000. Yeah, totally.
53:08
All these seed rounds for the AI companies are like 18 million or
53:14
20 million. That's a seed round.
53:16
That sounds more like he. Yeah, I mean I think counter to
53:22
what I just said about that you're, you know, the AI stuff
53:26
has been kind of insulated and it's kind of the bastion of
53:30
throw money at it. You know, like if you've got the
53:34
right pedigree and the right background and you're going to
53:36
go start, you know, something that's a kind of a journal like
53:40
an LLM company or general tool company, then yeah, you probably
53:43
got a lot of venture. But what's the, what's the, how
53:46
long are those going to be here? You know, some of them might be
53:48
here for a while, that some of them have got billions of
53:51
dollars raised, you know. But some of these other ones, I
53:55
think they'll just have a hard time competing with the ones
53:58
that are already light years ahead, you know?
54:00
Yeah, it is fascinating to me too to see that there's all
54:04
these different use cases. Like, I am a big fan of
54:07
following what happens in ecommerce and ecommerce.
54:10
Used to you used to be able to raise money for ecommerce and
54:13
now it's kind of like, do you really need to raise money for
54:17
this? Is it better just to get debt?
54:19
Like what? Yeah, what's the reason?
54:22
Is it gonna be one of those 100 or 1000 X outcomes?
54:26
I don't know. So it's hard.
54:30
But Nick, man, I appreciate you coming on here.
54:33
I've taken up a tonne of your time and this is super, super
54:36
cool for anybody that wants to reach out.
54:39
We'll leave all your details in the description so they can hit
54:43
you up on LinkedIn. And I'm going to keep using
54:46
gifts. I'm just going to plant my flag
54:49
in the sand right now, just spite what my daughters tell me
54:55
and my cousins that say I'm not cool anymore.
54:58
I don't care. I will take it.
55:00
I really enjoy the way that gives make me smile when I see
55:05
them. So I love what you are doing and
55:09
thanks for coming on here. I played with this to be true.
55:11
Thanks for having me. Thanks for all you do.
55:13
Take care. Hey Lazlo here.
55:18
If you're serious about Emma Loops, you hit subscribe right
55:21
now.