Latest / Agentic Conversations (formally mlops.community) / Founding, Funding, and the Future of MLOps // Mihail Eric // #200

Transcript

Founding, Funding, and the Future of MLOps // Mihail Eric // #200 · Agentic Conversations (formally mlops.community)

0:06
If they see you man, I know it's been awhile.
0:09
Ohh. How You have been a while.
0:10
It has been a while. How's it going?
0:12
I'm great, dude. I have been jamming to my new
0:16
favourite Spotify playlist all day and I'll give you two
0:22
guesses what you think it is. Let's just go with names.
0:26
Like like actual artist, Like like the like the title of the
0:29
playlist or the artist. Title of the playlist or genre.
0:32
I'll give you that too, if you're anywhere close to the
0:35
genre. I'm gonna say like like yoga
0:36
meditation. Like Zen?
0:39
I mean, I do listen to that a lot, but that's not what I've
0:42
been jamming to. You are kind of in the ballpark
0:46
with that. You got another guess.
0:49
OK, OK, so I was I was actually going, if that was wrong, I was
0:51
going to the exact opposite end of the spectrum and be like
0:54
heavy metal death punk, like Metallica.
0:59
But since we're in the in this, I'm gonna say something like
1:02
smooth jazz. Ohh, Miles Davis is a big fan of
1:06
Bill Evans. But that's not today.
1:08
Today you wanna know what it is? What's the what is it today?
1:11
Gregorian Chant. OK, you know that is not
1:17
surprising to me at all. That is not surprising.
1:20
It's like more melodic slow hum. Dude, it helps me get work done,
1:26
man. Like, it is so useful.
1:28
And here's why. I'll I'll say that you were in
1:32
the building when you said the yoga and meditation.
1:34
Because it's not just Gregorian chants.
1:36
It's Gregorian chants at 432 Hertz.
1:40
You know how they do those where they're like ohh this is special
1:44
heart vibration or frequency. And stuff.
1:48
Right, right, right, right. So it's like In Sync with your
1:50
like circadian rhythm or something like that or what?
1:52
Exactly. Just gets those Theta brain
1:55
waves going, or the alpha brain waves.
1:57
I don't even know what brain waves.
1:58
I won't go in, but it gets them going.
2:00
My favourite pseudoscience. She Exactly.
2:07
Exactly. Well, I don't know anything
2:09
about the human body. It's like I'm reading this book
2:11
about, like, longevity and how to, like, live a longer,
2:13
healthier life. And it's just, it's so unclear.
2:17
Like everything is kind of like hypothesis.
2:19
And I think like a lot of nutritionists make hypotheses.
2:21
You know, there's like 17 different kinds of diets you can
2:23
do. They're like the best thing for
2:24
you, from keto to paleo to to cardio.
2:29
Again, so I'm here's the ones that I subscribed to.
2:32
Then later you can tell me yours.
2:34
I do the intermittent fasting and I'm vegetarian but pretty
2:38
much vegan. OK, OK, so you don't eat till
2:41
like 2:00 PM or something like that.
2:43
I'll eat. My window is usually from like
2:45
12:00 to 6:00. 12:00 to 6:00. Unless I do like a sauna in the
2:52
afternoon night and then it's kind of hard not to eat.
2:55
Yeah, do that. Yeah, yeah.
2:57
But yeah, in general, whenever I eat at night, man, I just feel
3:00
like crap in the morning. So, so this actually like you,
3:03
You again, it's hard to tell whether it's a placebo or not,
3:06
but you actually, I know. Yeah, yeah, exactly.
3:09
I have no idea because I've been doing it for so long now.
3:11
It's been like three or four years that I've been doing this
3:13
for. So I have no idea if it's just
3:16
like when I eat at night, I'm like ohh, I shouldn't be eating
3:20
ohh but it's so good and then I feel guilty and my body does
3:24
that Total placebo or if it's actually like my body's
3:27
circadian rhythm, right says hey you shouldn't do.
3:30
This. Yeah, you just being
3:31
self-righteous. That's what's going on here.
3:33
You're one of those people that smells their own farts, you know
3:35
what I mean? Exactly.
3:36
And I I analyse it with another pseudoscience.
3:39
Those are the it's that's exactly what it is.
3:43
But dude, do you know what? This time of the year it might
3:47
be one of my favourite times of the year because the leaves are
3:50
changing colours. I know you don't get this in
3:52
your privileged position right outside of Stamford or wherever
3:58
the hell you are in California. You don't get any leaves
4:01
changing colours, do you? No, no.
4:03
So all kinds of leaves changing colours.
4:05
It's getting cold enough for me to be like, all right, cool,
4:09
There's a change here happening and you start to feel that.
4:13
And I get to wear the absurd sweater collection that I have
4:19
that I have vested in tears knowing that this one's my
4:23
wife's mom, handmade like in the 60s or 70s and nice.
4:29
Last time I was back and we were visiting them, she was like, ohh
4:34
yeah, I got all these sweaters that I made for my brother and
4:36
then he died. And now I have them again.
4:40
Do you want them? I'm like, ohh jackpot.
4:42
Here we go. They're so.
4:44
Warm. They are so warm, dude.
4:46
Yeah, Like like wool sweaters or something like Yep.
4:49
Yeah, and mid back in the day like with a lot of love and care
4:53
and. So yes, yes, that adds an extra
4:55
layer of warmth for. Sure, Yeah, totally, dude.
4:59
I appreciate this. I appreciate this.
5:00
I I was I I'd spent a few years on the East Coast where you get
5:04
the I mean like look as beautiful as a California is and
5:08
it's like year long you just have the best weather always I
5:11
will say when I was in the East Coast and you have those brutal
5:14
winters like brutal brutal winters that ends edition from a
5:18
winter to a spring when like for the first time like that there
5:21
is like nothing that compares to that like there's not a single
5:24
day in like California is like a 7 or 8 all the time that's like
5:27
a nine and a half you know I mean.
5:28
That's off the Richter scale for sure.
5:31
You like not only you are happy, but everywhere you go everyone's
5:35
happy and so it just adds to the collective vibe.
5:38
That's right. That's right.
5:39
Yeah, it's amazing. It's amazing.
5:41
I feel that and then flowers are coming out.
5:43
And flowers are coming out. Oh my God leaders.
5:45
Most of the thaw. And then you like strike.
5:48
Things start to bloom. You're like, oh man, this is
5:50
after that. Maybe it's the delta.
5:52
Maybe it's going from like wow, that was so painful and painful
5:55
and freezing cold to like wow, I just feel amazing right now.
5:59
But it's it's it's nothing compares.
6:01
Nothing comparison for sure. Yeah, alright dude.
6:03
So this is the catch up episode. I feel like it's been too long
6:06
since we chatted and you've been on here many a times as a cohost
6:12
and then you went to start. Can we give people the
6:15
background? Like yeah, you were focusing on
6:18
starting something in the NLP space before the chat should be
6:22
right, right, Right. Yeah, I can do something around
6:24
data quality and labelling. And then you went to try and
6:27
raise and the VCs were like, ohh, we don't want anything to
6:30
do with that. I think I remember you quoting
6:33
me one time that you met with 60 VC's and nobody said yes.
6:40
That was so rough time, yeah, for sure.
6:42
Yeah, I can imagine that would get your spirits down a little
6:48
bit and you would start wondering, is it me or is it the
6:51
idea and you are fighter and you said it is the idea.
6:55
It's not. That's right.
6:57
That's that's the way to avoid Eddie Eddie sort of
6:59
introspection. Just being like no, I'm
7:01
definitely not the problem. Exactly.
7:07
So then you pivoted and and it was right around the time that I
7:12
went and I was working for stability AI and you were like,
7:17
hmm, I think I can do something in the image generation space.
7:21
Yeah. So our paths continued to
7:24
intertwine in that light and we ended up doing stuff.
7:31
And I think we both have since pivoted away from, well, I know
7:34
I have. I'm pretty sure you have pivoted
7:36
away from doing anything with stability.
7:40
Not image generation though. You're still doing stuff there,
7:42
but right. There's all these questions that
7:44
I wanted to explore when it comes to building a business in
7:46
the image generation space. We talk a lot on the podcast
7:50
recently about LM and so this is for all those people who are
7:54
like, fuck me, another LMS podcast.
7:57
I don't want to hear it. Don't worry folks, we've got you
8:01
covered. Cause my man Michael is here to
8:03
talk about the image generation and just give us the background
8:08
on what you've been up to. Yeah, and and ohh, yeah, I
8:11
should probably tell folks like that didn't hear you or don't
8:13
know about you. You were before you went and
8:16
tried to start the NLP thing. You're working at Amazon on
8:22
Amazon Alexa. Yeah, you were.
8:24
From what? You went to Stanford, graduated
8:27
Stanford. So you've got the credentials,
8:29
which makes me think that when it comes to raising money, it
8:34
would have been easy if idea was good enough.
8:38
So I'm saying it's always the idea.
8:39
It's never it's never happened at all.
8:41
It was not. No, no, I I I've facetious.
8:44
Of course, of course it's. Even.
8:46
And I think you even had me on the deck or something as an
8:50
advisor. And so it should have been a
8:53
home run man. Yeah.
8:54
So yeah, must have been. That's right.
8:57
That's been a really bad idea. I no, but give me the give me
9:02
the breakdown. Like, what have you done since?
9:04
Yeah, yeah, for sure, For sure. Yeah.
9:06
So I mean I think the the first part of the intro is very on
9:09
point. This was Emma Lobs play, Data
9:13
play. Maybe a tad too late, honestly.
9:17
And you know, I'm sure these days you talk a lot about LM
9:20
OPS, but there was a predecessor to this ML OPS, and I'm sure we
9:24
can, we can wax philosophical and what exactly is the
9:26
difference between the two? Are they the same?
9:29
What's really changed, if anything at all?
9:31
But this was largely an ML OPS play before LM OPS in data
9:36
focusing on structured data. And there was, I think it was a
9:39
little too late. I mean it just got to the point
9:40
where there were too many players in the space the day.
9:43
It probably wasn't differentiated enough at the
9:45
time. I, even though I still a lot of
9:47
like the hypothesis that I had about data and data quality, I
9:51
think still will hold true and still will play out in the long
9:53
run. It's just like what?
9:55
Well, primarily that like I think a lot of it still has to
9:58
go end to end, right. That's in in the sense that you
10:02
have to kind of start from like, you know I was starting like
10:04
labelling, like trying to solve the labelling problem and then
10:06
expect that once you solve the labelling and you build out
10:09
through the rest of the stack, you'll start to hit into issues
10:12
around increasing the quality of the data through a training run
10:17
through serving. When you start getting
10:18
predictions from models. And then how do you kind of pipe
10:21
that all back into maybe even the labelling getting improved.
10:24
So you it's like this virtuous cycle where just because you
10:26
started the annotation, you go all the way through train and
10:29
you go through serving and then you get predictions, you kind of
10:31
need to bring all that information back in.
10:33
And I suspect a lot of that will still play out.
10:35
And in a sense this is, this is like how a lot of ML companies
10:39
have sort of evolved their offering and you know they start
10:41
with this like 1 niche and they all are kind of having to go end
10:45
to end because they want to own as much of that value stack,
10:48
that value chain that they can. And there you have to go either
10:52
left or right depending on where you start right and also play
10:55
out. Yeah, it's you have to wedge
10:57
yourself in there somehow and then try and you know market
11:01
share. That's just start reading more.
11:03
And some people, I mean, I think some companies have made it
11:05
easier for themselves and some of them made it harder.
11:07
I mean, at the time that I was doing, I was like, you gotta
11:09
start at the beginning, you gotta start labelling.
11:11
You saw labelling and then and then you're good.
11:13
And to be honest, like a lot of the VCs at the time were like,
11:16
you know, doing a labelling company in 20, what was it,
11:19
2022? It's like it's a tough sell
11:21
right now. It's really like a tough sell.
11:22
There's just too many of these players out here and you really
11:25
need to come in with something drastic.
11:28
Not only were there too many, but there was too many that
11:31
weren't making money. Right, right.
11:33
It was, it was tough to differentiate and so kind of
11:36
went back to the drawing board. This was around the time that
11:39
stability had actually just released their first model right
11:43
the first like open source version.
11:45
I mean Dali it's been around like he had been around but the
11:48
idea of an open source image Gen model had not yet been around
11:51
and it was at this time that I also met my now cofounder.
11:54
Want to give a shout out to Julia.
11:55
Super awesome person. Definitely one of the most
11:57
talented engineers and just thinkers I've ever had the
12:00
pleasure of working with. Super great.
12:02
How did you meet her Actually through a mutual friend.
12:04
So we were like we were both kind of iterating on ideas in
12:08
broadly LM's and and kind of Jenny.
12:11
I know her background is also technologist, but she was at
12:14
Google Research for the greater part of a decade and was just
12:19
someone that had, you know, really done a lot of really
12:21
great work and had built. She actually was on the Bert
12:24
team, right. So the Bert.
12:26
Bert, which is like the predecessor to all this LGBT
12:28
workhorse. Yeah, exactly.
12:30
The workhorse like Transformers Galore backward before
12:33
Transformers were a thing as a she really build out her her
12:36
career there doing great work and was very fortunate to kind
12:40
of be exploring at the same time that I was exploring.
12:43
And so we you know we're technologists we we validate
12:46
relationships by building things with people as I think you
12:48
should in general. I mean I think it's just
12:50
generally when you're when people come to like Ohh I'm
12:52
co-founder dating. I think you have to spend time
12:54
to build something with someone even if it's not both coders.
12:58
You just have to you have to come into the conflict basically
13:01
you have to like hey we're we have deadlines we have things
13:04
that we need to make and meet at certain time frames And so we
13:08
did that right. For technologists, it's easy to
13:09
just say we're going to pick and build something.
13:12
And so we both had experience in kind of multimodal applications
13:16
at our last jobs. Like I'd even at Alexa, though
13:18
I'd spoke with a lot on dialogue systems, like pure textual
13:21
dialogue. A lot of what we were thinking
13:22
about as well is how we can combine like text and other
13:26
modalities like image. You know, we were thinking a lot
13:28
about how can we put Alexa on like haptic devices and like
13:33
robots and things like that. And so we put that, you know, we
13:36
were and she somewhere in the background had a similar
13:38
background looking at a multimodal stuff at Google.
13:41
And so, well, let's try to build something that combines visual
13:44
and textual generative AI. And this was like a little bit
13:47
of a hack that we did this, like a stunt.
13:50
We made this site called Rick and Mortify and Rick and Mortify
13:53
Dimitri. You might remember this was a
13:55
little like kind of toy thing that we were like, hey, we both
13:57
like Rick and Morty. Let's see if we can do something
14:00
with this, you know, kind of piggybacking off of our love for
14:03
the show. We made this application where
14:05
you could kind of pick up the characters that you want in, you
14:08
know, your favourite characters in the show.
14:09
You'd provide this little like textual premise for what you
14:12
want a new episode of the show to look like.
14:14
And then using generative AI in the background, both kind of
14:17
LM's like traditional textual LMS and visual LMS. We created
14:20
almost like a comic book of sorts where we generated the
14:23
plots of a new episode including the dialogue.
14:26
And also you know for each episode or like little segment
14:30
of of that new episode, we would, we would actually create
14:33
a visual of of that of that what that would look like in terms of
14:36
the actual like kind of image itself.
14:38
And we put that out there and you know visual Jenny, I was
14:41
still kind of a thing people were excited about this and
14:43
people love Rick and Morty. So it kind of blew up like we
14:46
were getting like like 10s of thousands of people using it in
14:49
a very short amount of time. Like holy crap this is, this is
14:51
interesting like what's maybe we should kind of pull on this
14:53
thread what what getting distribution in that way enables
14:58
you to do. And this is just I think a
15:00
common problem in startup is like distribution is always an
15:02
issue. When you get distribution like
15:03
that, you get reached out to and you have a lot of interesting
15:05
conversations. One of the conversations that
15:08
fell out of that project is we actually were able to speak to
15:12
director, a screenwriter who was like hey, you guys are doing
15:16
this thing storyboarding. Seems something could be
15:19
interesting. And so we we, we kind of went
15:21
down this rabbit hole of trying to build like a an AI powered
15:23
storyboarding tool which used a lot of the same technology and
15:27
we did that, we did it for a number of months and that from
15:31
that we spawned off kind of a a next succession of products that
15:35
that we're working on right now. But yeah, that really took us
15:38
down. It was a little bit piggybacking
15:40
off of what we thought was interesting set of problems
15:42
technically combined with just following the rabbit hole of
15:45
like opportunity, right, Where something shows up, you're like,
15:48
well, I don't know what a storyboard is, but let's figure
15:50
out what we can do with this. And we went back and forth with
15:53
this person who's our advisor. So it was just happenstance
15:56
meets capitalising opportunity. There were so many things that I
16:01
wanted to ask on this. Please cannot.
16:04
So can I just say just at the beginning of what you were
16:06
talking about, founder dating? That is something that is so
16:10
Silicon Valley, like San Francisco.
16:14
I'm not sure if there are many other places in the world where
16:18
it's just totally a normal thing to be founder dating.
16:22
And you have these things that you go through and you say, hey,
16:26
you know what, let's build something together because we're
16:29
founder dating and you might be my match.
16:32
Exactly. That's right, left, left swipe,
16:34
right swipe or whatever it is. Yeah, I'm I'm surprised nobody's
16:38
made like a Tinder for founder dating.
16:40
You know what it is I think the YC has a like cofounder matching
16:43
service, which is closest thing I've seen.
16:45
They don't have an app though. It's like a web app.
16:47
There's not a mobile app, but it's pretty close.
16:50
It approximates the similar field.
16:52
Yeah, that does feel like it. I mean, let's go into this idea
16:56
of like building a business off of something because just for
17:03
those, for the lay of the land and especially when you were
17:06
starting this business, yeah, there was stable diffusion and
17:09
it was probably when stable diffusion like 1.5 was out and
17:14
that was the, yeah, that was like the upgrade from the first
17:17
one. And it was when things were
17:19
getting really good and there was a lot of hype around it.
17:22
And then stable, the Fusion 2.0 came out and that's one option
17:26
that you have. The other option is Dolly, which
17:30
is alright, but not really that good.
17:32
And I don't even know if they have an API.
17:34
I don't even know what that looks like and how you can
17:36
interact with Dolly, but it definitely didn't.
17:39
You're as good as stable diffusion 1.5 and then you have
17:43
mid journey, which is basically the one that takes the cake as
17:47
far as quality goes, right? But it's so hard to work with
17:50
mid journey because every time that I've tried to work with it
17:54
you have to go into Discord and as far as I know, like did they
17:58
just open up an API or is that still you have to work in
18:02
Discord? And as far as I know, it's still
18:04
in Discord, you know? Oh my God.
18:06
And well, kudos to them because the products got way better
18:10
because of that, because they're getting all that feedback and
18:13
they're getting to know what people like.
18:15
And that's why the Mid Journey Image Generator right now is so
18:23
superior to stable diffusion in every way, right?
18:28
I honestly think that the latest versions of like SXL I think has
18:31
really closed that gap like whatever has been done between
18:35
and maybe it's just the fact that the model about
18:37
significantly larger and more data higher quality data.
18:40
I I suspect that's probably a big part of it.
18:42
It is not. It's not a it's not like
18:45
majority is blowing them out of the water to give a really
18:47
honest. OK.
18:48
So, so it's it's feasible to build a business with state
18:53
stable diffusion now? I think so, I think, yeah.
18:56
The prompts prompting is easier because I know that there's a
19:00
tonne of resources when it comes to mid journey and prompting and
19:06
being able to get solid prompts and stable diffusion.
19:09
There was a lot for 1.5 and you could also kind of get the same
19:16
stuff when in the beginning, yeah, when it was like stable
19:19
diffusion 1.5 and mid journey and then later on it felt like
19:23
prompting got way different for both of them and you would have
19:27
to have totally different prompting guides and ideas on
19:31
the prompts. And so now what are you doing?
19:34
Like how are you? Are you hosting your own stable
19:38
diffusion on your own cluster and what does that look like?
19:41
Yeah, yeah. I mean, look, I I like mad props
19:45
to to these companies for building out.
19:47
Yeah, stability and and Co. The journey is a little bit more
19:50
of a walled garden. But I mean they really just hit
19:53
the kick off a huge influx of interest in these things for a
19:57
while and and I think that was that really did push the field
20:01
forward a lot. What I will say is it's, I don't
20:05
think it's enough. It's like you can't just use a
20:07
vanilla SD model off the shelf and expect to have a business
20:11
that that is certainly not true. And one of our, our realisations
20:15
was that in spite of all the, I mean it feels like years have
20:21
passed, but really it's been like a year since stable, stable
20:24
diffusion came out there. One of our big hypothesis and as
20:29
a company is that the ecosystem around image generation,
20:33
generative AI and building applications around that is
20:36
still very underdeveloped, right.
20:38
So in a sense, it's almost like it almost feels like a little
20:41
bit like an OPS play where you're recognising that in order
20:44
to enable a greater swath of developers and companies to
20:48
build out Imagen stuff, it's still not enough, right?
20:51
Like even though we have these, these suites of API's, I mean
20:54
really just from stability primarily.
20:56
And Dolly, yes, there's also an API, but you know, you have to
20:59
build out something that's a little more, a little more
21:01
extensive. Yeah, so you can't build just a
21:04
vanilla stable diffusion model and it feels like the tools
21:08
aren't there and then you have the what you need is more
21:11
developer tools, but besides these like APIs that you have or
21:15
a suite of APIs because with mid journey you still don't have
21:19
APIs which is crazy to me. But yeah.
21:21
Stable Diffusion or Stability has APIs that you can use.
21:26
Yeah. And I think one of our
21:27
realisations in in building out by starting to build out like a
21:31
verticalized application of generative AI, which is what the
21:33
storyboarding thing was. The which is where we started
21:35
and where we've bowled from is that in spite of, yes, there
21:39
being a suite of APIs out there like you know you can leverage
21:42
like a stability like the the your mileage may vary, right.
21:47
And you have a lot less control over the kinds of models that
21:50
you want to include. And you're not going to be able
21:52
to do it with just an off the shelf stable diffusion model,
21:56
right. Like you're going to be able to
21:57
build an entire business off of that.
21:58
You're going to have to fine tune things.
21:59
You're going to have to leverage Laurae's tremendous amount of
22:03
Laura's and just to get that style that you want.
22:05
So there's still a lot of adaptation that needs to happen.
22:08
And once you have to adapt things, once you break out of
22:11
that, that quick start shell they give you, it turns out that
22:17
that the ecosystem is not making it easy to get started and to do
22:20
all that customization you want to do.
22:22
And so like our hypothesis here is that there is one of the
22:26
first things we started with is like, hey, you have an image
22:28
generated like maybe you want to fix things up a little bit
22:30
because they're not perfect, right.
22:32
Like and that's there's like a last mile problem sorts that is
22:35
actually critical when you deal with different kinds of
22:37
verticals. They're not just going to take
22:39
something that you generate from the journey or stable diffusion
22:42
and take it as it is, right. They want to modify it a little
22:45
bit. Maybe it's close, but not
22:47
exactly close. And so you know all about this
22:49
is that you really have to build out more of this ecosystem to
22:53
enable things to be built and and primarily a developer
22:57
ecosystem because all right, but this is also that you will,
23:00
you'll be able to like the future is all about building
23:02
these multimodal applications that will you need to provide
23:05
something like a stripe for visual generative AI and that's
23:09
a little bit where we're going. Wait, stride for visual.
23:13
I don't know if I understand that, but it sounds sexy as
23:16
hell, man. Tell me more.
23:18
I mean it. I add Unicorn with another
23:21
buzzword. How does it not sound?
23:22
Amazing, no? Dude, where are these BC?
23:27
They're throwing money at you now, man.
23:29
No, let me let me illustrate. I mean, you know the I think
23:33
that that little phrase is really just meant to try and
23:37
clarify, but it ended up confusing.
23:39
I mean when you think of like Stripe when it came out and like
23:46
there's I guess they gotta work on the analogy but but here's
23:49
here's here's what I'll say here's what I was like that like
23:52
Stripe when it came out back in like I think it was like ohh
23:55
nine maybe 2010 they're they're big big big value add was like
23:59
you wanna enable payments to be done in in the on the Internet
24:03
right. Any person who's building up
24:05
Bootstrap business, indie hacker, what have you, wants to
24:08
build payments. But it's really hard to deal
24:09
with like payment API's. Like it's just really, really
24:12
hard to do so. And so Stripe made it like
24:15
delightful to build payment into an application because they give
24:18
you like, here's like here's 4 lines, put it in there and it's
24:21
just going to work. You know what I mean?
24:23
And to have that kind of a guarantee in something like
24:27
Visual Geneva, it just doesn't exist today.
24:29
It just it won't work. It won't be like, here's 4 lines
24:31
and it just will work out-of-the-box.
24:33
It'll be like, hey, here's maybe 4 lines and you it might give
24:36
you 400 error where the server is not working and maybe the
24:39
formatting of the thing isn't right.
24:40
And good luck. If your image looks the way you
24:42
want it to. And so there's still a lot more
24:44
finagling that needs to happen to get something to the point
24:46
where it's it's exactly kind of what you have in your head.
24:49
So that's we want to get to the same level.
24:51
Like developer delight that Stripe offered.
24:53
OK. I see something that's.
24:56
That's like Jenny. I like.
24:57
Visual Yeah. You're talking like, yeah, the
24:59
developer experience. That's right.
25:01
That's right. That's right.
25:03
I get it. And one thing that I was
25:05
thinking about when you were talking about how the APIs that
25:08
are there, yeah, stability is offering stuff, but it may not
25:13
be what you want. And then the models that you can
25:15
use aren't exactly what you need.
25:17
And especially if you want to fine tune them right and you
25:20
want to use their API, then it gets even more complicated.
25:23
And have you looked into all these other basically API or
25:29
model providers as a service type?
25:31
Things like doesn't replicate or some of these places do that.
25:38
Yeah, yeah. I think that we're, we're
25:39
starting very pointed in our offering, right.
25:43
I mean the way we've done it is I think a lot of the they should
25:48
have maybe some business might fall into as they try and do
25:50
everything they try and spread themselves too thin.
25:52
You know, I mean they're like, hey, we're going to, we have to
25:54
support everything because our investors are knocking on our
25:57
door and we need like a we need like we need to offer a
26:00
gazillion things because we have $100 million.
26:02
You know, and that's just that's a little bit of the the that
26:06
that the siding way of your soul that you you subscribe to when
26:08
you when you raise a certain amount of money.
26:10
So what ends up happening in the long run is you get a lot of
26:13
breath, not a lot of depth. And So what we're trying to do
26:16
is start from just like a core suite of things that will work,
26:19
that we believe will provide value and then iteratively build
26:22
that in a way without really sacrificing and compromising the
26:25
developer experience. So what we've done is, yes,
26:27
we've released a suite of APIs around some core bits of
26:30
functionality. We've done a lot of image
26:32
editing stuff. We have some stuff in the
26:34
pipeline right now around video again which we will iterate
26:37
iteratively and very slowly deliver the roll out.
26:40
But we're also doing that in tandem with showing people
26:42
what's possible. If you have these things, what
26:45
can you do? So in addition to these APIs
26:47
that we've released, we've also built an applications.
26:50
We're like, hey, you can build out like a almost like an Adobe
26:52
Express or they'll be Firefly basically very easily, you know,
26:56
I mean like you have these APIs and you can make an editor in
26:59
browser editor that does exactly that.
27:00
Or we did. Once we had the APIs done, I was
27:03
able to put a Discord bot almost you know that could literally be
27:06
in the same channel as like at the mid journey.
27:08
Bought in a matter of like a few hours where I was like hey you
27:12
you can make a call to mid journey and then write it for
27:14
you can call our discord bot and you can iterate and fix up your
27:17
images and do other things that maybe you can't do in mid
27:19
journey. So having a good bit of
27:23
foundational scaffolding really enables you to build out
27:26
applications so much faster right?
27:28
And that's that's really the the bet here and and what we believe
27:31
needs to happen. Because at the end of the day,
27:35
what you see a lot of when it comes to the open source
27:38
community around stable diffusion, I feel like are those
27:42
UI's, but you don't necessarily necessarily see it as many back
27:48
end pieces for the developers, right?
27:51
And I think it was a lot of catering to that artist who was
27:56
playing around on their computer and got some Google Collab GPUs
28:02
for a limited amount of time. And then they're like, whoa,
28:05
cool, I could do so much with this, right, Right.
28:08
And then you get that UI and you're cool.
28:11
And I also there was one that I really liked for a while.
28:15
I think it was called like Prompt Hunter or something like
28:19
that. Or Prompt Hunt, maybe like
28:21
Product Hunt, but Prompt Hunt? I can't remember the name
28:25
exactly, but for a while you could swap out the different
28:30
models that you were using. You could say like OK now I want
28:33
to try this with yeah mid journey.
28:35
I want to try this with stable Diffusion and it would give you
28:38
clues and help on defining the prompts and it would also give
28:43
you different like tags so if you wanted to add stylistic
28:47
features. Right, right, right, right.
28:49
And that's incredible as the front end.
28:51
But eventually they stopped offering Mid Journey because I
28:56
think they probably just ran into the same problem that
28:58
everybody ran into. Like Yep, having Mid Journey
29:01
work outside of Discord is really hard.
29:03
And So what does that look like? How does that work?
29:06
And so now what you're doing is, if I understand it correctly,
29:10
you were like, alright cool, I'm going to go on the back end.
29:12
I'm going to host my own stable diffusion.
29:14
I'm I get the feeling you're using some AW.
29:18
I mean I I have a very strong soft spot, strong soft spot,
29:23
strong. So I have a soft spot for a WS
29:25
of course and and love them. But you know we also in a sense
29:29
we can be agnostic, right. It's not something where you
29:31
have to be. I I think a lot of businesses
29:34
nowadays try to be multi cloud anyway because it gives you a
29:37
lot of flexibility. And and you're just offering the
29:42
API and the UI and then you're going upstream and being like
29:46
here's some cool stuff that you can build with this.
29:47
So you're yeah, yeah, yeah. Full stack.
29:51
That's right. That's right.
29:51
Exactly. It's like, hey, here's something
29:53
you could here's something that's powered by these building
29:56
blocks. Like here's some building
29:57
blocks, something that's powered by these building blocks.
29:59
You might like that as well. But it doesn't have to be the
30:01
only, you don't have to restrict your imagination to just a thing
30:04
you run in your browser. You can, you can go and build a
30:06
Discord bot. You can go, you can go build a
30:09
print on demand service, You can go build a a what have you.
30:12
I mean, there's so many applications that you can just
30:14
leverage go go build a video game editor, you know, I mean
30:16
like there's just so many things that can fall out of this.
30:19
So it's it's about giving those blocks and then maybe painting a
30:22
little bit of the narrative around what that looks like.
30:25
What are some of the other things that you have thought
30:28
about building in the different niche specific?
30:33
Verticals, yeah. Yeah.
30:34
I mean a lot of it is. It comes out of conversations
30:37
that that we get from customers and potential customers right
30:41
where we're where. We couldn't imagine some of the
30:43
things that fall out of as use cases like it's just something
30:47
that I would never even imagined.
30:48
For example, one of the ones that came out, it was very
30:51
unusual but also very cool is when someone was like, hey, I
30:55
build a print on demand service. And what that affected what
30:58
they're doing is they to their consumer, to their user.
31:02
It's like a platform, it's a, it's a UI where you can go and
31:05
you can sort of like iterate on some image or some pattern that
31:08
you want. And then from that pattern they
31:11
will really allow you to generate wrapping paper and you
31:14
can then get like custom delivered, you know, made and
31:16
custom delivered to you. So you'll be like, hey, I want a
31:18
picture Demetrius face, you know, with like a Christmas
31:21
sweater and make that like the Envelopes community, like
31:25
Christmas. Wrapping do that, give me the
31:27
URL I'm going to do this. I would do it and I expect that
31:30
would make a lot of money. It's gonna be our swag for the
31:34
holidays. That's right, that's right,
31:37
Dimitrio asked. Have you been a good ML OPS
31:39
engineer? You know if you've been naughty
31:42
or nice, that is for sure. So, so you know they that's the
31:46
service they're offering. But for them they were like hey
31:48
we you know we're maybe like a little bit light on the AI
31:51
development front. But these APIs seem like they
31:53
could really make our job of filling out those images much
31:56
easier, cleaning them up or or kind of iterating on them.
31:59
And then you can imagine that in a service like that you're just
32:02
like our IP is just plugging into their pipeline, right into
32:05
their workflow because the the, you know the product is what
32:08
they have. But the that's just like
32:09
something that I would never even imagine.
32:11
And and we're getting like inbound from people that are
32:13
like what I never thought that was a problem, but but that's
32:17
cool. I mean like adding watermarks in
32:19
in some some use cases or people that are being circulated like
32:22
housing in SF where they're trying to like extrapolate kind
32:25
of the looks of of of like of new housing projects.
32:29
Like one thing I also want to call out is and maybe you know,
32:33
we'll touch on this as well. It's just like this idea that
32:35
Jenny is so hot right now. It's like Zoolander all over
32:38
again, right? Like it is so hot.
32:40
And like people are kind of moving from one thing to the
32:44
other, right? Like it did start with, in a
32:47
sense stability was like a thing.
32:49
And then Chad GP sort of came out and then we had like Lama 2
32:51
and so now like texture. LM's like all the all the rage
32:54
but that there's still like an ecosystem that has to be built
32:58
out to really enable to really like in a sense democratise a
33:02
certain kind of technology. And a year that did not happen
33:05
like in the year that in this in like the six months that people
33:08
cared about stable diffusion and then moved on to llama two and
33:11
and you know all this technology stuff like we didn't get that
33:14
ecosystem to really reap the benefits of what we have.
33:17
Like if you look at a lot of the tools that people use for
33:20
interacting with state-of-the-art visual Gen AI
33:23
models just a few that come to mind they'll just name drop here
33:25
like automatic one-on-one is like an open source project.
33:28
Comfy UI is like another open source project.
33:30
I mean these are, these are great open source efforts.
33:32
I really don't want to think like they've done huge, you
33:34
know, really helped the community in a lot of ways.
33:36
But when you look at something a little more production ready or
33:39
stable, maybe a little more user friendly, that's that can't be
33:43
where we stop, right. They can't be where we stop in
33:46
terms of offering ways of interacting in ecosystem.
33:49
That was we would run into those issues.
33:51
Like I don't really know how to use this, like it's still really
33:53
complicated. Like you think Discord is
33:54
complicated like man, using some of these can be like really
33:57
tricky where you're just like watching YouTube tutorials for
34:00
like hours. I think this is how you do it,
34:02
so it it feels like there has to be a better solution, right?
34:06
Yeah, it's so funny because it is a lot of YouTube tutorials
34:10
when you're going down that rabbit hole and like, have this
34:14
image of how I want it to look in my head.
34:16
And I feel like I could get it there, but I don't know what to
34:20
do to make it get there. And I remember at stability,
34:24
there were like the prompt experts and dare I say prompt
34:29
engineers, right? They and I would ping them every
34:31
once in a while and be like, here's the photos that I'm
34:33
getting. I want the tree to be inside the
34:37
box, not like the box next to the tree.
34:39
And here's the prompt I'm using. What do I have to do?
34:42
And it really felt like they were just like, well, I'm, I'm
34:47
gonna say that you probably should try and put tree inside
34:52
the word box or box inside the word tree or put box and tree
34:56
together. And you're like, what's this is
34:58
not scientific at all. You're just throwing shit at the
35:00
wall. Like, I what?
35:03
I will say a lot of times that shit worked and I was like, how
35:07
how did they know that? So or they're like put a hyphen
35:12
between box and tree and you're like what?
35:14
No way that's gonna work. And then you try it, it's like,
35:16
Oh my God, that actually that worked.
35:20
So yeah, there's there's some funny things.
35:22
It's like hyperparameter. Tuning of of Yeah.
35:24
Of yesteryear. You know or I mean like Once
35:27
Upon a time people would be like you.
35:28
You gotta use the alpha learning rate of .001.
35:33
You're like what? Really?
35:35
Because otherwise your loss is like this.
35:36
And then you said .001 like it was like 1 decimal point over.
35:39
You're like huh? That's just.
35:41
That just works exactly the way I wanted to.
35:43
That's crazy. Yeah, and it's very tribal
35:46
knowledge to get there. That's right.
35:48
You wouldn't know that unless you were working with someone
35:51
who could help you and you could ping and ask.
35:53
Or you go into some community and you get lucky.
35:58
That's right. The person who knows stumbles
36:01
upon your question. There's a lot of that.
36:03
And so I can see that pain. And I can also see that
36:10
potentially what happened was, is the height moved on from the
36:14
image generation tools. And so the ones who were left in
36:18
the image generation world were people like yourself.
36:21
And you're trying to figure it out and you you feel like, Oh
36:24
well, I've got a pretty solid use case here, right?
36:27
I got people that are interested.
36:28
I'm doing something verticalized, right?
36:31
Right. Then you hit a wall and you
36:33
realise maybe I need to go down a few levels and start building
36:39
out the pipes before I can build the house.
36:43
You know, like the foundation is not there yet.
36:45
And so let's try and make that foundation and then we can build
36:48
on top. Of it.
36:49
I think that's exactly right. That's correct.
36:51
Yeah. Yeah, yeah.
36:52
It is funny though, because I in the LM space, I feel like you've
36:56
seen a lot of people try and jump at building out the
37:00
foundations. Because yes, it also feels like
37:03
there's a lot more things that you need to do and areas that
37:09
need to be improved. You know, like whether it's
37:12
latency or whether it's truthfulness or trustworthiness
37:17
explain ability or whatever it may be.
37:19
And I also think that the the beginning of the hype wave, you
37:24
had so many people saying like ohh, we could use it for
37:28
everything now, right? It's for everything you whatever
37:32
you can dream up, you can do it with AI now.
37:36
And basically now I'm. I feel like that's calmed down a
37:41
bit and we're starting to see, OK, there's these use cases that
37:44
it works incredibly well with, and then there's other stuff
37:47
where it was a little wishful thinking, right?
37:50
Right. It's a good point And and look,
37:53
I, you know I've been an AI longer than this hype cycle,
37:57
right? I mean I I got into this like
37:59
well over a decade ago at this point and so this is not like
38:01
the first hype cycle I've seen in in a I like when I went into
38:06
I mean it was maybe like the first new modern hype cycle like
38:09
2012 or so when deep learning really became a thing again.
38:13
And then maybe 20/14/2015 when maybe like machine translation
38:17
systems really started getting good and people were able to
38:19
translate languages which was a big step forward in speech
38:22
recognition also are really good.
38:23
And then maybe like 2017, there's also kind of another
38:26
little wave that that was happening.
38:27
So iteratively people, you know, it, it gets better, it gets
38:31
bigger in terms of the wave. I would say this one though, the
38:34
last like one to two years is like completely unlike anything
38:38
I've seen. You know it's.
38:39
Like when I think. Largely, a lot of it did come
38:42
down to very easy to use consumer facing applications of
38:47
AI. Right.
38:47
When you have something like a GPT or like a Dollie where any
38:51
layman can be like, have you heard of this chat GTP thing?
38:54
You know what I mean? It it felt like when when back
38:56
in the day, like I remember this one and it was like a little
38:59
really about a story The first time I was at a Thanksgiving
39:02
dinner in like 2017 and random family friends are like here
39:05
there's like Bitcoin thing. You're like, yeah.
39:07
I was going to say that yeah. If if like, if they're asking
39:10
about Bitcoin, like, you know, there is like a huge some bubble
39:14
of sorts that's happened here. And so I, you know, my, my, like
39:18
maybe this little bit of hot take like I think there's going
39:20
to be a correction in terms of the hype.
39:22
And that's exactly what you've suggested.
39:23
Right. Is that that it went from
39:25
everything is possible to like, look a lot of things are gonna
39:29
be great. Not maybe everything but if you
39:32
like start to ask yourself like the very deliberate question of
39:34
where is this really gonna shine.
39:36
There's awesome stuff to build. Maybe not everything but awesome
39:41
stuff to build. I think that that getting to
39:44
that second secondary question of what's the awesome stuff to
39:47
build is maybe the contraction a little bit from the.
39:50
Yeah, I don't want to put this person on blast and say the name
39:56
of the the Twitter influencer that said that, but I just
40:00
realised we were at peak hype when someone said.
40:04
On Twitter, something like I realise I'm in a cafe living in
40:08
the future and look around. Nobody understands the effect
40:12
that AI is gonna have on our lives right now.
40:16
It's like, dude, I mean, you're gonna be able to summarise some
40:22
shit and ask questions through your notion database a little
40:25
bit better, but I'm not sure we're talking like you're living
40:30
in the future. Yeah, yes, for sure.
40:35
And who knows what it's going to be in two years.
40:37
Hopefully the amount of tooling and the amount of energy and the
40:41
amount of building that is happening in this space will
40:44
make a difference. But when I see that shit, I'm
40:47
just like, Ohh man, I gotta, I either gotta go outrageous on
40:55
Twitter and start just following, drink the Kool-aid
40:57
and start saying shit like that, or I gotta get off Twitter and
41:01
save my mental. Health.
41:02
Yeah you gotta get new friends. It's it's a really good point
41:06
too. I mean we could probably spend
41:08
another hour talking about that but but this this site height
41:11
brings the rise of like new influencers, you know, I mean
41:14
that's always happened every new hype cycle there's like people
41:17
that like Ohh man like I should start talking about this and
41:21
then you have these people that have like never done anything in
41:23
AI that are now like better now the go to sources for learning
41:27
about AI and. I'm just that one.
41:30
I totally missed that one. That was that would that should
41:33
have been me. I slept on my laurels.
41:36
I'm the guy who knows nothing about AI.
41:37
And then boom, I just saw the wave come and crashed right in
41:41
front of me. No, but you know it's it's it's
41:43
like it's a you have to learn to play that game.
41:45
It's I think it's and and it's like, do you?
41:47
Want to play that game? Exactly.
41:49
Is that something that you Because these people are like
41:51
they they they'll have three like 3 * a day until and they
41:54
have to live on some of these social media platforms.
41:57
And. Yeah like like I I don't wanna
42:00
dig them completely because it's good to spread information as
42:04
long as it's the right information.
42:05
It's a little bit as as like I don't call myself like OG you
42:08
know I mean but like someone that has been in the field a
42:09
little bit longer I it is a little.
42:12
Like like because I don't want to call myself that, but I am.
42:18
Then I mean, here's what, like, I I I have, I have written
42:21
papers longer than this, right? And like I was writing papers
42:23
about this stuff before the fact it it's it's more that I'm I, I
42:27
I just wanna make sure that that when people are.
42:30
I don't like people that know talking about to speak with so
42:32
much compact so much confidence as much confidence like the
42:36
until you like like just at least don't blow things out of
42:40
water. Like provide people nuance.
42:42
Like there are people that have become influencers where I'm
42:44
like that's still legit. You're giving the nuanced
42:46
version of the answer and you're not taking people off this weird
42:51
hype train. We're like everyone is just like
42:53
this is amazing. Yeah.
42:54
This is amazing. This is so amazing.
42:55
Right. Yeah.
42:55
I know. This is amazing.
42:56
Right. And you're just like there's
42:58
like these people talking and just being like, this is great,
43:00
right. Yeah.
43:00
This is totally great. You know.
43:01
I mean it's it's a little weird. Yeah.
43:04
I feel like it's going to hurt the field.
43:05
And Lauren, be honest. Yeah, cause it's gonna set the
43:09
expectations so high and then when it takes another right five
43:14
years to get anywhere, I think even if where we are in five
43:20
years is incredible and it's huge breakthroughs, right?
43:23
Right. If it's not at the spot where
43:26
it's like AI's completely taken over and we don't even have to
43:29
work anymore, right? Right then people are gonna be
43:34
disappointed because. We've seen this, we've seen this
43:36
and this chat box. This is Chat Bots 2017.
43:39
This is exactly what happened. People were like, Oh my God,
43:41
chat everything and and and then.
43:44
Which they suck. It's that hedge cycle, man.
43:46
It's literally goes, it goes up and then it goes down and then
43:49
and then it kind of like gradually goes back up.
43:51
Like, I still think that conversational AI.
43:53
I mean, obviously I'm biassed, but I still believe
43:55
conversational AI has the potential to really help a lot
43:58
of fast society. But when you have too much
44:01
happening, too much excitement happening too quickly, people
44:03
yeah, get very inflated expectations.
44:05
And then they get disappointed when in reality the science is
44:08
still like all this Gen AI stuff is amazing, but there's still
44:12
hard research questions that need to be solved.
44:14
Yeah, it's not completely commoditized.
44:17
There's still a lot of things that researchers need to figure
44:20
out to make this work. Exactly.
44:22
And unit economics. Unit economics, A lot of money
44:26
being lost right now. Represented by the big players
44:30
100%. But also I think one thing that
44:33
I wanted to talk to you about a little bit, because you played
44:35
in the Hollywood space and I know you have like 1 foot in LA
44:40
and one foot in San Francisco. Yeah, yeah.
44:42
I think the hype also brings a bit of a negative aspect in that
44:47
regard because people don't know what to believe and so they
44:51
expect that AI is going to take their job and then you get
44:53
writer strikes or you get whatever and you don't recognise
44:59
where it's at necessarily unless you really deep dive into it.
45:03
Because if you just ask a few questions to ChatGPT, you could
45:07
get the impression that whoa, shit, like my job's gone
45:10
tomorrow, right? And so being in, like when you
45:16
were hanging out in Hollywood, would you get a lot of people
45:21
talking about how they were afraid or they would not like
45:25
the idea of this whole storyboarding tool with AI
45:29
because it's going to put someone out of a job, the
45:31
storyboarder out of a job or whatever.
45:33
Yeah, for sure. For sure.
45:34
Yeah. I mean, I was actually literally
45:36
on a panel like 2 weeks ago at a Film Festival in LA.
45:39
So it was funny because the way the panel was constructed was it
45:43
was like the AI people on the right side of the table,
45:46
moderator in the middle, and then like, Hollywood folks,
45:50
producer, actor, who was in, like, marble movies on the left.
45:53
And it was just like us talking at each other.
45:56
And man, were they coming at us. Like, they were coming at us
46:00
like, they were like, what are we gonna do?
46:02
Like, what's going on? And this is, you know, this is
46:04
like eight or nine months and, you know, like many months into
46:06
the strike, right. I mean, and now it's thankfully
46:09
come to a resolution of sorts. But yeah, I mean, look, even
46:12
even this many months in, people are still like, we don't
46:14
understand what's going on here. It feels like, what am I going
46:17
to tell my, my, my crew when like, all of a sudden, like, you
46:20
know, they're they can't send their kids to private school or
46:22
whatever, You know, it's things like that.
46:24
So like it just, and I don't mean to belittle the problem,
46:27
actually feel like, you know, they they they're totally
46:29
justified to feel concerned. I will say though that as you
46:33
said, it's like about hype makes people just just how quickly
46:38
they just get scared by things that are changing too fast.
46:42
You know, like like change has to come at like an organic pace.
46:46
It can't come just like if it comes too quickly, people get
46:48
scared. And that's why I think I saw
46:50
happening definitely in Hollywood was just a very
46:53
traditional industry where it's been very much the same, you
46:58
know, like the biggest. They got shaken up by Netflix
47:00
right a few years back and that was like them blowing their
47:03
minds like, what do you mean Netflix?
47:04
Like what's you know, what do you mean the streaming thing
47:06
like and that that really turned them on their heads.
47:09
That flavour of industry getting hit by tragedy, BT getting hit
47:13
by Dali, getting hit by you know mid journey, all these things
47:15
like whoa whoa whoa, we don't move this quickly.
47:18
You know. I mean we're not we we don't, we
47:20
don't we can't accommodate that kind of change so quickly.
47:23
But that being said, I think so yeah, it was a big sticking
47:26
point in all these strikes, right.
47:27
I mean there was a thing they were all trying to get figured
47:30
out and I think a lot of it just comes out to misinformation.
47:34
I really comes down to like education and and just getting
47:37
credit and attribution in a way that people get compensated for
47:39
their work. I think that just has to that
47:41
piece. Once that piece is figured out,
47:43
I I really think that the issues aren't as big as people blowing
47:47
them up to be. And then as you said as well
47:49
literally during this panel people are like, but AI is
47:52
coming and I'm like you guys don't understand how far AI is
47:54
from like all this stuff that you're saying it will do like
47:57
this is like decades, like in my opinion, right?
48:00
Like this is like to get to that level it's gonna take a while.
48:03
So it's a lot of, there's like complex issues here play, but I
48:06
think that the you can understand why they're coming.
48:10
And you did say something interesting there.
48:13
When it comes to the attribution, yeah, that's a
48:16
fascinating 1. The other thing that it sparks
48:18
in my mind is about videos itself and how we had ChatGPT,
48:24
which went for the scriptwriters or whatever.
48:27
But it's pretty obvious that a script written by Chachi EBT is
48:32
probably not going to get greenlit for a film anytime
48:36
soon. Or if it is, it's like tongue in
48:39
cheek right type thing, right? You can brainstorm.
48:43
If you're a script writer, it's probably a great tool to help
48:47
you and your productivity. But it's not like, oh, write me
48:51
a script about XYZ and now boom, we've got Hollywood blockbuster.
48:56
At least in my mind. Maybe I'm wrong.
48:58
No. Yeah, that is way over hyped in
49:03
in my mind. And then you have the image
49:05
generation, and there's all kinds of ways that you can use
49:09
that within the greater film creation process, right?
49:15
I do see how it's like, whoa, we just got hit on both sides of
49:20
the coin with now it's the writers, then it's the editors
49:23
or it is the creative directors in a way.
49:27
And then if you add on video on top of that, I can see why
49:32
people would be like, Oh my God, it's coming.
49:35
And it's coming for everything, every piece of this.
49:38
But I also feel like video is really far off as much as I do
49:42
see do it. Like Runway Mill is doing some
49:45
incredible stuff. Yep, Yep.
49:47
With or just what it enables you to be able to do in video and
49:54
with AI is absolutely amazing and mind-blowing and but I still
50:02
don't think it's anywhere near like, OK, now we don't need any
50:08
visual effects people on our set.
50:10
Right. No.
50:11
No, Definitely not. Definitely not.
50:13
And and I mean, you know, we could, we could go off on like
50:16
how the world could play out like even if even even in the
50:20
years, many years away when that is a reality, like what does
50:22
that really mean for how content would be created, right.
50:27
And I, you know, I have some hypotheses, I don't think it's
50:29
going to be, you know, people will still be a huge part of
50:31
this process fundamentally. I mean for me, and this is even
50:35
what I said to to to these folks, you know, at at the
50:40
festival was like, look, you have to believe that there's
50:41
something like fundamentally human about the arts.
50:44
You have to believe that if it's not true, then you will be
50:47
replaced by AI because there's nothing fundamentally human
50:50
about it. But you have to really like in
50:51
in your heart believe that there is something that cannot be
50:54
replaced like a human, a sentiment, a human memory, like
50:58
a human experience. And that's something that we
51:01
have to imbue our systems with or like, you know, how to guide
51:04
them into maybe the resistively help us play out a story or or
51:09
you know, kind of get to the point of of acting out a story.
51:11
But but that has to be something that's lived in a more human
51:14
fashion. So, you know, I still do, I
51:16
still do subscribe to that. And as you said, I mean
51:19
technology is it's it's quite a ways away.
51:21
I mean, yes, it's cool to have 1015 second clips that show some
51:26
level of emotion and something that's still very underdeveloped
51:29
and it's a hard, hard problem, right.
51:31
I I really suspect there's going to be more smaller segments of
51:36
the of the puzzle of the value chain where we can start
51:39
integrating some of these tools. It'll just make you more
51:40
productive. I mean that's that's really the
51:42
goal here. Always like AI is supposed to be
51:44
assistive, not really supposed to replace these things or these
51:47
people. Yeah.
51:48
And so that's your whole vision right now is how do we build the
51:51
building blocks to help that value chain.
51:55
And if I'm understanding it correctly, you're trying to
51:59
build something that is generalizable.
52:01
So it's not only in the Hollywood vertical anymore.
52:06
It's like if you're building something for the Hollywood
52:09
vertical, you could build with your tool as the basis, but it
52:13
doesn't have to be in that vertical.
52:15
Yeah, Yeah, it's something more horizontal.
52:17
I mean, you know, there are certainly a few verticals where
52:20
we, we suspect and we've seen a lot of residents, I mean, video
52:23
games for sure, entertainment obviously, you know, interior
52:28
design, architecture, marketing. I mean, there'll be ones where
52:30
it comes a little more naturally.
52:32
But again, we're not excluding ourselves because we get these
52:35
things that show up at our doorstep like print on demand
52:39
like you know housing project development and and and and kind
52:42
of construction. So just things that you really
52:45
just didn't even anticipate. And so it is a little more like
52:48
a horizontal layer for sure. And what do you think about
52:52
Milos Verse LLM OS, Do you? There's a new stack, right?
52:57
And there's also prompting in the mix.
52:59
And there is the evaluation. Yeah, it's LM's, yeah.
53:04
There's probably also the iteration if it's image Gen like
53:10
where where's your head at with respect to that?
53:13
I actually think that there was a, there was a post I think
53:16
Eugene Eugene Yuan put out on Twitter that I actually quite
53:20
liked where he took, remember that old, I think his D Scully
53:25
paper from Google which is like, you know the old school
53:28
Godfather paper where it's like, hey, here's all the here's all
53:32
the stuff around ML code, right. Yeah, here's the technical debt.
53:35
There's not the model. Yep, Yep, exactly.
53:39
And there's like 15 other boxes around, just like a little tiny
53:42
thing. Little box that is the mill
53:43
code. And he had replaced that very
53:45
tongue in cheek with like LM code.
53:48
And I thought that was a good way to to, to synthesise what I
53:53
also agree with, which is look at the end of the day all this
53:57
stuff like, yes, yes, some of these models are bigger so that
54:00
there is like challenges, maybe different envelopes challenges
54:03
around or different infrastructure challenges around
54:06
how do you host a 15 billion, 70 billion parameter model Like,
54:10
yes, that is hard. And that is not something that
54:12
when we were doing logistic regression, you know, random
54:15
forests 2 years ago and that's how people were operating.
54:19
So yes, it is a different infrastructure challenge that
54:21
will be solved. So that that is a slight
54:23
difference. That being said, I think that a
54:26
lot of the same concerns apply. Like a lot of the exact same
54:29
ones apply. I mean data still ends up being
54:31
the name of the game. Do you read any of these papers
54:34
coming out for how you're getting improved systems like
54:36
Dolly Three and and any new Genii system like the the the
54:41
novelty is still the data. They're like, we're getting
54:43
better data, getting better captions.
54:44
We're getting better this and that.
54:46
And that's what people are still figuring out.
54:48
So in a way I feel like we've had some differences and we
54:52
progressed maybe a little bit, but do we really need like a
54:55
whole new LM OPS things like new frame, you know, phrase sounds
55:01
like it's other VC sort of oriented thing.
55:03
Totally. And and I would rather just say
55:06
like it's it's ML OPS with like now there's a new model that we
55:08
have to support. You know what I mean?
55:10
And and here's the interesting thing is like like if you see a
55:13
lot of these ML OPS companies of the last two, you know, before
55:18
LM became a thing, who were around for 2-3 years who are
55:21
pivoting very easily into LM OPS.
55:23
Like they're just like, it's not, it's not like they had to
55:25
like, holy crap, we need to go raise like another Brazilian.
55:28
Thing. Our whole strategy.
55:29
Yeah, Product needs to. Absolutely not.
55:32
Like, I was talking to Piero from Party base, you know, a
55:34
good friend and they've been doing Ludwig for, for years,
55:37
right, Uber, it says the open source project.
55:40
And then they built up, you know, this build business that
55:42
was just traditional ML stuff. And I'm like, how is this
55:45
changing anything for you? Like really not at all.
55:47
I mean maybe a little bit of the paper on the landing page, but
55:50
we have the building blocks have already built out.
55:53
So you know what I mean. So we're just we just have more
55:55
inbound now from people that are like I don't know I what does
55:59
this LLM stuff and so they're getting a lot more customer
56:01
inbounds, but fundamentally the technology is like exactly the
56:04
same you know. I mean 100% that is awesome to
56:07
think about. The other piece I I was going to
56:10
mention something, but you got me off track when you mentioned
56:13
Piero and I was thinking like, Oh yeah, they just you can pivot
56:16
and say now we do fine tuning instead of something else.
56:20
Even everybody's got that story now and so it's funny about
56:25
that. And when it comes to that Eugene
56:29
Yan post, I think D Scully actually made a new blog post.
56:35
It did and that's what it was. That's what Eugene was
56:37
referencing And then in that in that and so dude, incredible.
56:44
Always a pleasure. I gotta go to another meeting.
56:47
Sadly, we got to do this more. Hopefully we'll have you on soon
56:51
enough to give us an update on how things are going.
56:54
And if anyone wants to use the tool, we'll leave a link or you
56:59
can tell us the URL right now. Yeah, lab dot storia dot AI.
57:04
Or if you the our company's story AI.
57:06
If you want to you can also reach out to me directly just my
57:10
name's Michael at storia dot AI. Or you can just reach out to
57:13
founders at story dot AI. We're we're happy to talk and
57:16
learn about use case and build. Cool stuff.
57:19
Boom. There we go dude.
57:22
Alright man, thanks for doing. This.
57:24
Hey man, thanks for writing me. Really enjoyed it.