Join us as we delve into the fascinating world of Artificial Intelligence (AI) by interviewing the brightest minds and exploring cutting-edge projects. From innovative ideas to groundbreaking individuals, we're here to uncover the latest developments and thought-provoking discussions in the AI space.
This idea was kind of born out of Matt and I just
having random AI conversations, but some of the
stuff that he's built, like, hey, it wasn't
intended to be a pet product. It was messing around
with some things and a pet and a product came out of
the other end. But also, you know, Matt can go on
about this at great length about how everything we
built is not about pets, but just about scaling,
training, and the entire infrastructure in the
cloud at AA. The most reasonably economic way you
can do it, would you say that's around hair, Matt?
Yeah, so Schematical is, like he said, my
consulting company. I help businesses that are
getting traction and trying to scale up on AWS. So a
lot of times they just try and boot up bigger
servers, but you can re-engineer it to be more cost
effective as you scale up. So with that said, I was
digging into AI. I wanted to do some experiments
with stable diffusion. I had my YouTube channel
where I was, I'd done video game stuff before, so I
was going to generate game art, you know, pixel
art, using stable diffusion. And at the same time,
I was like, well, how do I host this on Amazon in a
cost effective manner? And there's some really
expensive ways to do it, which I didn't like. So I
derived a different way that only turns on when
people are using the program and shuts off
afterwards, which saves me a lot of money and lets
me host this. And then from there, Dominic and I
were showing him some of my results and he's like,
let's do a pet thing. And from there, it seems
people are reaching out to us outside of Drawing by
AI to utilize this technology to host their own AI
stuff and do their own inferences or generate
images on Amazon in a cost effective manner as
well. And I've open sourced most of it. So if your
viewers are at all interested in that, Terraform
scripts are online. So like for the Drawing by AI,
that works only when someone's using it pretty
much from what you're saying. Like it's only
turned on like on EC2 or like in Amazon or how does it
work actually? The expensive part. So the so the
cheap part, just taking web requests, you know,
you sign up for a user account, but you haven't
uploaded your pictures or anything like that yet.
That stuff's on Lambda and that spins down
whenever no one's using it as well. So it's
basically $0 there. And then the expensive stuff
with the GPUs, that boots up only when we've got
work to do. So training a model, generating an
inference, and then it shuts back down. Got it. So
are you like for Drawing by AI, is it are you using
stable diffusion to produce images? Yeah,
dreambooth. All right. So yes, taking stable
diffusion, we're in it. We're taking, there's the
base stable diffusion that can be trained with
dreambooth or other methods to produce these
intermediate models. We are using an
intermediate model then to train it again on
specific dogs to get the artistic output of that.
Got it. So the intermediate model has been trained
on tens of thousands of pieces of art from all over
the world. So there's several trainings involved
with the process, but that intermediate one, we
get that problem on In Face. It's one of the more
common ones that's out there. So can you just tell
us a little bit maybe about how you guys met? You
said you met in Wisconsin. How did you guys meet?
How did you get involved in AI in the first place? So
we met years ago. I think it's probably been 12
years now. By various tech, there was a good tech
meetup group in various tech communities in
Madison, then Dominic hired me to work for him once
or twice. I hired him to work for me once or twice.
And we were actually just kind of fresh off a
project together. And yeah, so we wanted to come up
with something. As far as AI goes, I've
experimented with my YouTube channel, a lot of
neat neural evolution of augmented topologies.
And I wired that into Minecraft and I started doing
that somewhere around 2019. Interesting stuff.
Wasn't really monetizable. So then I switched
over to stable diffusion, done some LLM stuff,
some fundamental stuff. I'm no AI expert or ML
expert. I'll tell you that. I'm an expert at
hosting stuff online. So what I've done basically
is I try and bridge the gap between the people that
have models and they're like, hey, I know this, we
can go to market with this. I've got an AI trained to
do this or PhD and they're like, but how do we make
that available to 10,000, a million people? And
that's kind of where I come into bridge that gap
there. So you're like the, are you like the sort of
like an architect or engineer to say, to show
people what the most cost effective way to host
things on like Amazon is for the most part? Yeah, my
title would be web application architect. So
you've got a lot of, got to have a lot of different
overlapping skill sets to make that happen. Like I
said, ML is not my super strong point, but teamed up
with the right people. We can do some pretty good
stuff. And Schematical is it's a company, it's a
consulting company, but it wasn't started for AI,
it sounds like it was started more for helping
people with their architecture, their tech
architecture. It started as a product, which I'm
almost embarrassed to admit what the product was
years ago. And then it became synonymous with, I
had to consult because the product didn't take
off. And it just became synonymous with my name
online at this point. So in reality, there is a
company Schematical consulting, but I'm also
known as Schematical just online when tagged a lot
of times. So how about you, Dominic, how did you get
into AI? I'm a software developer by trade, but I
spent better part of the last 10 plus years on the
like the management side of technology. And I
really wanted to swing my career back towards
tech. So about a year ago, I completed a data
science bootcamp and the content I thought was top
notch actually excellent, excellent content.
And that's where Matt and I started talking back
and forth about AI ideas and used to be live jam. So
they just kind of all of our interest kind of just
intersected at about the same time and place. Very
cool. So just tell me a little bit about this
project, draw AI, right? I mean, we've seen it
before. We looked at it. What is the project? How
did you come up with it? Why? It's a little bit about
its origin. I was working on retooling my skill set
to make sure I could service my machine learning AI
based clients. And I figured there was, you know,
something out there we could take a swing at. And so
I was generating actually put photos of my dog in
and myself, my mother and a handful and train these
fine tune them so they would be able to generate our
face. So I could say, I want a picture of me on the
beach in a, you know, in watercolor or something
like that. And it could do that. But the really cool
thing is what it did with my dog. It actually did
some really amazing photos there. And I was
sharing that with Dominic. And this is where I'll
pass the story to him. Yeah. And I said, look at
that. I'm like, well, somebody's willing to pay
for that. And, you know, that was a very rough model
at that point. And we had to take it and obviously
build the web infrastructure and the business
model around it. But we looked at doing a lot of
different things, you know, should we do people?
Should we do dogs? Should we do cats? Should we do
and just we started with dogs. It seemed to make a
lot of sense. People you first of all, the, the,
the, the attention to detail on a person, the
quality you want to get on a person is a different.
You're going to look at that far more critically.
And we wanted to avoid any sort of people making
pictures of people that weren't that sort of risk.
So dogs seemed like a really good place to start out
and obviously got your attention from a sea of AI
things. So it must have looked interesting to you
as well. Yeah. It's a pretty cool project. I think
the people really love their dogs. Anything that
people can do for their dogs. I mean, everything
you see is like, has a dog version to it. And
anything you can do with that, I think draws the
attention of people for sure. There's a very large
social media audience on Instagram, on TikTok.
Absolutely. So was it, was it Matt's idea? Matt,
was this a, it sounds like this was originally your
idea and then you sort of worked together with
Dominic on it or how does it? It's tough to claim it
as my idea. I mean, and my idea was to host stable
diffusion in the cloud and basically to make an
easy way for people to train models. Cause I hadn't
seen other people like other products where
people could train their own models. You can use
open AI to do an image to image transition. So you
upload one image and say, change it, but you
couldn't upload like 50 pictures of your dog,
yourself, whatever, and then have it generate
completely new artworks. So that was my part of
this. And then Dominic was the one that said,
let's, let's focus in on a project for pets, you
know, drawn by dot AI, we can reuse that name for any
number of things. But for now let's launch with
pets. So Dom and Dom has a background and he did
launch a pet product at one point previous to this.
So I want to give a nod to that. And that's a fun
story. If you want to. Please tell us about it,
Dominic. For a couple of months during the, the,
the first Donald Trump, the, the, the election
that Donald Trump won. I sold Donald Trump and
Hillary Clinton dog poop bags on Amazon so that you
could put, yeah, whichever, whichever party
effect you could buy the opposite of one and pick up
your dog's poo with it. If you do a little research,
you'll, you'll find me on Univision, a two minute
spot on Univision in New York when this was going
on. Hey, you're kidding. They're both crowds.
This wasn't our first idea. We've been bouncing
ideas. We've actually been having live streamed
idea jams every couple of weeks for a few months. So
it was a process of just bouncing things around to
see what really makes sense. I will say you brought
the focus. That's for sure. So do you have dogs of
your own? I've got two and they're a handful.
Dominic, I think you travel too much for one right
now, but you're thinking about getting one. I know
that I've had several and I've got one that should
theoretically be arriving sometime in the
spring. So give us like an overview. I mean,
obviously I've looked at the tool before, but give
the audience an overview. What does it look like
when they go on to drawn by drawn by dot AI? What does
it look like? How does it work? And what do they need
to do? So the first thing you have to do is add a
single or more than one picture of your pet. And
this is kind of the funnel in action, if you will. So
you put in one or more pictures of your pets at which
time that you then have to create an account
username, password, things like that. And at that
time, we're going to ask you to upload at least 10
pictures of your pet altogether because really to
learn the animal from different angles. So one or
two pictures is not going to give particularly
good results. In fact, for best results, you want
to upload 15 or 20 high quality pictures of your
dog. From that, Matt, I think this is where your
front end AI takes over. You could talk more about
that than I can. We've had problems with people
uploading spam images. So I actually wired in
YOLO, which is the object recognition algorithm
to verify it. And if it doesn't find X amount of
photos, it'll flag it. So that's kind of fun. And
then we, those are all uploaded because uploading
is cheap. It's training that's expensive,
actually. And so those are uploaded. And if
they're not flagged, then they go right into the
queue. Otherwise, one of us will review it and, you
know, give them, give, you know, approve, deny
based on that. So then we train a model using that
those photos you've got. And we, each model is
unique. So there's every time you do one of these
things, there's about a four gigabyte model that
gets saved to the cloud that we then can reuse to
generate your photos on the fly. So we generate, we
call inferences. What do we got? Six or eight of
them right now, Dominic? Nine. I forget. Nine?
Nine. So we've got nine. So that's your example.
Your dog is a captain of a ship. Your dog is a
watercolor painting. Your dog is XYZ. And if
you're interested in signing up with us, you can
download the HD version of that or the underwater
market versions, I should say. And HD version
coming soon. And we can additionally, if you want
to keep, you know, with the subscription, we'll
keep sending you more of these. So eventually you
can get a Halloween version or, you know, holiday
version of it or. And so you keep getting these fun
little pictures. So that's pretty much how it
works. So are there choices on the website like you
want a Halloween picture or you want a pirate
picture? Or do you say that in a prompt? How does
that work? So we looked at all of our competitors
and they did two things very, very differently.
One of them was they would make you pay before you
could see a single output. Whereas we're letting
people see the images first. But the other is you
would have to choose, you know, they've got 50
different styles. Choose 10, choose 6, choose 8.
Creating the images is relatively inexpensive
compared to training the model. Once you have the
trained model, that's I'm not going to say a bulk of
the cost. That's a large percentage of the overall
cost. So we decided to give everybody everything
and take the decisions out of it. Now, what you see
is when you see your dog as an astronaut, you mean
you've played with AI enough. I'm sure whether
it's the language models or the image generators
that you can't just say, give me something as
simple as that and expect a really deep textured
response from it. So that astronaut is about 25 or
30 words to create the astronaut image with an
interstellar background and all of these other
things. So there's a lot of tools out there that
give users the box here. Type what you want into
this box. But unless you've really spent time with
these prompts to understand how to add the texture
in the depth to the prompt to get back a
satisfactory or a really nice result, you're
going to be disappointed in it. So going forward, I
would love to be able to offer the users much more of
a hand in the creative process. But again, I can't
just give them that box because it's going to
underwhelm in the results. It doesn't mean there
couldn't be a set of descriptors and scenes and
props. For example, I wanted a picture of my dog in
front of the Eiffel Tower wearing a beret and
watercolor where you could check those major
descriptors that add that. And but behind the
scenes, then the Eiffel Tower will be a bird's eye
view of the Eiffel Tower in spring with the tree. So
there will be that word will be behind the actual
button that the user wants to click to add that to an
image. A really important point to point out to the
audience out there is like you might go to their
website and be like, well, I can just do this on
mid-journey. I can just use this to do this myself
on stable diffusion. But like I use mid-journey
every single day. And like it sometimes it takes 20
different prompts to get to what you want. So these
guys are taking the pain of that out for you. Like
they're giving you a fine grained final image,
which sounds like that's what your
differentiator is to other platforms. Yeah, our
average user isn't a prompt engineer. You know, my
mother, for example, wouldn't even know where to
begin with it. But she's got photos of dogs. She
wants to run through it. And so that's the beauty
part of what drawn by does in the front end there.
Now, I will say the underlying engine behind it is
perfectly capable of taking in whatever prompts
you want. So we have like other people that don't
want other businesses, they want to build their
own stable diffusion product. And they want
complete autonomy and flexibility to just throw
in prompts and train their own models without
having to worry about any of the fancy stuff of
hosting it. And we're actually we're totally
capable of supporting that too. You know, they
just make a simple restful request and they get
whatever they want back. And that's through
Schematical you're saying in that in that
instance. That's yep. That's it's through
Schematical. It's it's part of the technically
the underlying engines called chaos pixel,
because I'm very bad at naming things. But it's
it's kind of in layers is that Schematical is a
consulting company that built it. You know,
there's drawn by and then underneath it, the
engine that's hidden that no one sees is chaos
pixel. And is this the are you guys working on all
this sounds like you have ideas out there. Are you
working on any other projects right now other than
drawn by AI in the AI space or I'm not looking at
anything specific and so far as projects. I'm I'm
pursuing some other things, but no, nothing else
like this. Yeah, I mean, just keep in mind those two
layered projects, which we're supporting,
supporting both of those and actively working on
selling both of those. I don't have any other AI
stuff, just my client work. So are you are you
planning on sticking you mentioned something
about the human aspect of it? Like if you're
straying away from the humans want to stick to the
dogs? Are you planning on sticking with the dogs
for now? Or do you see it possibly going to humans at
some point? I don't think we'd have a problem with
humans. I don't want to speak for Dominic here, but
we've got to master the user acquisition side of
things. If that makes sense. If there's a market
for it. Forever is a mighty long time, but it's
definitely not on the road map. We've had several
people ask us about different varieties of
animals. And I think there's a lot of products in
the person space that I've seen appear out there as
well. So there's people that take a picture and
say, give me one for LinkedIn, give me one for my
social media, give me one. Those tools already
exist in earnest. So when it looks at the dog, you
said you use YOLO, maybe that comes into play here.
I'm not sure. But when it looks at the dog, how does
it know that, you know, how does it verify this is a
picture of a dog versus maybe this is a picture of a
gaiting pig or something like that? Where is the
line drawn and how does that work? So to be clear,
the YOLO is just to filter out people that upload
pictures of their shoe. So we primarily do that to
save money on the training part. YOLO is trained
with something like 30 or so objects that can
recognize the version I'm using. I can't remember
offhand. And so it's not incredibly thorough, but
it does a good job just detecting dog, not dog. It's
dog or 29 other things. And so if it doesn't detect
at least X percent dogs, each one of the photos has
at least one dog in it, then we send them to review.
So should we know that like this particular AI is
just a front-end JavaScript sort of tool. It's not
doing the training on the back end. So it's a gait
that it has to pass in order to get to the training
where the stable diffusion. And that's not to be
confused with the training. So the training, we
pass it two things. We pass it the instance photos.
So that's the photos of that specific dog. And we
also pass it a class of dog, which basically makes
it so it's not over-trained. So if I was training
human face, I'd take 10 pictures of my face and pass
that as the instance. And then I'd pass in about a
hundred other photos of other people that aren't
me. So it doesn't over-trained in my face. So we are
doing, we are passing in other dog photos when we do
the training as well. So if that makes sense. Yeah,
that's good. Yeah, I was just wondering because I
mean, dogs are very different. You have a
Chihuahua, you have a Wiener dog, and then you have
like a Great Dink or something like that. So
they're completely different when you're trying
to recognize it. It's like, well, if it's
recognizing a Wiener dog as a dog, it's
recognizing a Great Dink as a dog. Why not a Guinea
Pig or something like that? So that's what I was
trying to figure out. And it's not a particularly
sophisticated AI. It does label things often, but
it's good enough that it stops 80% of the, 90% of the
garbage images that people were uploading that
consumed actual resources to train. I will say
this, I probably shouldn't, but I will say
somebody got some rabbits through once and
actually believed they were dogs. That would be
interesting. There was something else that I
mean, jumping back to prompts for just a second.
There's something that it was really interesting
that we learned through this process that you can
write the greatest prompt in the world for a dog for
an image set. But you give that to 20 or 30 other
image sets of varying qualities and you get very,
very different outputs. So there was such an
iterative process to creating prompts that not
just worked really good for one type of dog or one
data set, but rendered consistently well across
dozens of different datasets of varying
qualities and varying types of dogs. And also,
there's certain types of dogs. Since you
mentioned there's this set of 700 and something
generic dog images to keep from overfitting. But
sometimes certain dogs, specifically I've
noticed all white dogs sometimes get a little bit
of more crossover into that generic dog class and
don't always come out fully white. And there's
other types of dogs that come through brilliantly
like Labrador Retrievers, for example, because
it's just a really common sort of size and shape of
dog that's really easy to identify. And it helps.
So it's two things, the prompts that you guys are
writing on the back end that nobody sees. And then
also the amount of pictures people upload. Like
the more pictures they upload, the better chances
the resulting images can be. I would say quality
over quantity. Quality over quantity. If you give
me 10 really good pictures of, you want to upload
pictures that positively influence the output.
And by that, I mean you want, you know, proud
standing pictures, not derpy laying on your back
photo. So if you give me three or four really good
headshots from a couple angles, two or three good
standing shots and two or three good sitting
shots, all from different angles on different
backgrounds that are well cropped and there's not
obstructions in them, you will get really, really
good output. And another thing to note is that you
can over train it. If you give it too few images or
say we only gave it four images and they were all
very similar, you're not going to get any
creativity with the poses there versus if you do,
you know, from a couple different, you know, good
angles there, you know, 20 or so. Now you're going
to get a little bit more variety in the poses you'll
see back. And when you, when you just go over again,
how the workflow works, when a user uploads all
these images of their dog and then hits submit or
whatever the button is, it will, what do you guys
spit out exactly? Do you spit out a bunch of
different pictures like examples of like, uh,
astronaut or a cowboy or whatever the the models
that you guys have are trained on? Yeah, if you go
through the drawn by AI process and you get your
photos back, those binary images that we return, I
think it's what six of them for nine different
inferences, that's everything that it exports.
Okay. And so right now it lets you do like what a
couple examples for free and then there's a paid
version. You can see all 54 of the output images
with watermarks. If you want to remove the
watermarks, then we, we ask that you pay for that.
And what is that? What's your pricing model look
like right now? Yeah, we're at a 20 US dollars right
now for that. And 20 US dollars unlimited
downloads for that month. So you've got it, you
know, and then, you know, if we end up doing more
holiday type stuff in the future, then you'd be
obligated to renew if you want more of that stuff.
That's very cool. Yeah, people, people will do
anything for their dogs for sure. I can see that.
And when I saw it, I thought it was, I mean, I just saw
some of the generic, you know, like examples that
you had and they were pretty awesome. I mean, I
could see people like printing them out, putting
them on their wall and stuff like pretty cool stuff
for sure. Physical products is also on the
roadmap. There's, there's some really good
international partner fulfillment partners out
there, but things like that. Coffee cups, socks,
blankets, stretch canvas. Yeah, I could see that.
I could see that for sure. I mean, like one of the
things that the problem is with AI is like, if
someone wants to go and play on mid-journey, I
mean, my parents don't even know how to copy and
paste on their computer. Like they have to go on
Discord, which is hard for even a developer
sometimes to learn how to use it. Like good luck,
you know? So I think there's a, there's like two
different, two different customers out there for
AI. People that just don't want to deal with a tech
background at all or any kind of tech and this fits
in a great niche for them. So it's easy to go on run by
AI and just, it's pretty simple compared to going
on some Discord channel and trying to figure it out
because it's not easy. I heard somebody who, who is
also in generative AI, a very different segment of
the space, but they explained it is their goal is to
be the application layer that sits between the
model and the user. user. So instead of just going
into a plain generic chat GPP box, you would add
descriptors and qualifiers outside of that box. I
want this to be warm. I want it to be like public
relations. I want this style. And then it would
apply those styles and that texture to the content
that was being generated. Building up that like
drawn by AI is the one that takes away all the
technical stuff. And then below that is Chaos
Pixel, which takes away any need to know Python or
any of that stuff. So there's different levels of
abstraction there, depending on how technical
you are. And what's again exactly Chaos Pixel?
That's the engine underneath it. So if you wanted
to run stable diffusion, but you didn't want to
ever touch Python code, you know, you could just,
you could basically with some API requests and you
know, basically like in hand you an SDK of some
type, you can upload the photos, then tell it to
train a model, then forever just say, Hey, I want to
this type of, you know, I want to generate, you
know, something else for this model. I want to, you
know, I want these images generated with it. So
it's someone that's technical, but doesn't want
to get into machine learning or worry about the
internals AWS or how their spend goes. It handles
all that form. Got it. And so it's not really a
publicly released product. Like we don't have
like web pages for it, but we're finding that there
is somewhat of a demand for that. One question I
have is when you're doing the model, can you
clarify exactly what that is the models to me, it
sounds like from what you're saying, just makes it
easy to make pictures of specific things, right?
Like an astronaut or a dog at the Eiffel Tower. So
you have a separate model for that. What's the
difference between creating a model like that and
then actually just putting a prompt in for a dog at
the Eiffel Tower? Like what's the difference? You
can think of the model as the machine and the prompt
is the input into machine. So that model is a
specifically trained model that knows your dog.
So that when you say, I want the picture of my dog in
this scenario, it already knows your dog and the
parent model, the stable diffusion base knows
about the Eiffel Tower and the other things in the
background already. So really, you need both in
order to achieve anything one without the other is
useless. And so as of the moment, places like mid
journey, don't let you custom train a model or fine
tune is the technical term, a model to your face or
your dog's face or another, you know, something
else you want to customize it to the best you can do
is maybe an image to image. And so you upload an
image and you tell it remove something, but it's
not actually like that intuitive. It's not like,
Hey, give me a picture of my dog on a beach with a fly
in a kite, you know, it can't and it wouldn't know
your dog's face from that one image. So it creates
that extra level of customization that right now
we're not seeing it a ton of big tools. Got it. That
makes a lot of sense because again, back to my
examples of using mid journey every day and trying
to do something very specific, it takes so long. So
it sounds like the the models take away all that
work for you because it's going to know the dog's
face exactly what it looks like. And it's going to
plug it into the image because it knows the size and
shape of the dog's face. Whereas if you're writing
a prompt for that or uploading an image, it's sort
of in a way guessing it's not trained on that right
away. So it takes a huge amount of the workflow out,
which you can save and use again, which is a huge
time saver. Yeah, there's no amount of text prompt
that you can feed mid journey that's going to make
it know what your dog actually looks like. Right.
Right. That makes sense. So something we like to
ask people in general is just where do you see AI
going as like AI in general, where do you see it
going in maybe five years, 10 years? Where's all
this headed? The question alone, I think is far too
broad to answer, right? Because AI is really
moving its way into everything. But in that same
context, like it's a bit of a paradigm shift. AI is
going to be incorporated in some way into darn near
almost everything. So your search results are
going to be AI. You know, the when you go to inbox and
you do a search, it's no longer going to be just a
text based search. It's going to be an artificial
intelligence search. It's going to more truly
understand what you're looking for. It's going to
be in customer service. It's going to be in quality
control. Is it going to drive millions of people
out of work and cause the task? In the short term,
no. But over long term, much like the mechanical
revolution, all of a sudden, you know, within 10
years, people didn't need to buy wagon wheels
anymore because of the automobile coming along,
things are going to shift. It's going to slim down
some areas when people learn how to use it to
augment productivity. So it's not going to
replace journalism, but could it replace half of
the proofreading department? Or it's not going to
replace your marketing and advertising team? But
is it going to augment it and let it run leaner and
smoother? Absolutely. But that's just a couple of
small areas. It's touching on every piece of
electronic interface we have is probably going to
be in some way influenced by artificial
intelligence in the next five or ten years. And I
totally agree with that. And I'm going to jump kind
of pivot just slightly. But right now, as far as a
prediction for kind of how the business
environment, the entrepreneurs chasing this
dream right now, as we're seeing a ton of people
throw a ton of money at this, anything with AI in the
name, you know, they're just throwing money at
that. And that leads me to think that there'll be
something of a bubble bursting at some point. And
the way to survive that is to do it leaner. So what
I'm seeing right now is people are like, okay, I
want to boot up this big server that'll run on a, you
know, on Amazon and be able to do all the workloads I
need. Oh, no, I'm spending an arm and a leg. Well,
I'm going to get a discount if I commit to a year out.
And so they do something like that. And they get 25%
discount because they committed to a year out. But
then who's using the product? You know, if you're
doing that before you validated, the product will
even sell. And so the only reason we can do what
we're doing is because we made it so it could scale
up and scale down. So anybody out there that's
thinking about chasing an entrepreneurial dream
with AI, don't make sure you can scale down as fast
as you can scale up and make sure that you're not
breaking the bank on it, because that's going to be
who weathers the storm at the end of the day. Sure.
Yeah, it's, I think there's definitely going to be
a bubble pop here maybe next year, or who knows
when, but it's going to happen at some point of the
other. Most of these people, like for example, we
write an AI newsletter, there's probably a
thousand of them out there. So that's what I think
about all the time, like how are we going to weather
the storm when things get bad? Because they
probably will at some point of the other. But
Schematicals are probably can help people like
out there that don't want to save the costs,
because if you go on Amazon and try to do yourself,
you're going to spend a ton of money. Just train it
yourself to figure out how to do it in the first
place. Or even if you're a developer, like I'm a
developer, but I hate going on Amazon and figuring
that kind of stuff out so Schematical could help
people like that as well. This is where there's an
entire field of ML Ops dedicated to just this and
CICD pipelines. And this really goes to show that
there is such a specialization between even more
so, I'd say, in data science machine learning. It
used to be the separation between the developer
and the infrastructure, right? And how to go from
one to the other. But even now, looking at the
machine learning pipeline, you've got four to six
different specialties that are completely
different skill sets along that pipeline from,
you know, on the theoretical side and the model
development side, you know, math, statistics,
linear algebra to, you know, to math's end of it,
the ML Ops to where it really touches the machine.
And you've got the whole application, machine
learning, training, feedback loop in there as
well, all different skill sets. I appreciate the
kind of shout out there. But I will say that I've
open sourced pretty much the entire like an entire
GPU running on batch, run pie torch, run dream
booth, all that stuff. I've open sourced into
Terraform scripts. So if you're not familiar with
Terraform, you basically can run Terraform apply
and it will boot that up inside your AWS account and
it will run and it's designed to scale down, like I
said, and scale back up. So that's out there for
free. If anybody wants it, it'll help you, you
know, rising tide, raise all ships. So I, you know,
if you have questions about it, come find me and I'm
on Discord all the time. But other than that, no,
I'm trying to offer something up to the people
because I am a firm believer that, you know, the
more people we can help, you know, not break the
bank with these ideas, the more interesting
innovations we're going to have in five years, you
know, but if you, if you say screw AI, because I
can't figure out how to afford it, then, then that
stifles, you know, all of our creativity. Yeah,
100%. Yeah, some people can get a bad taste in their
mouth if they try to start a business, spend
thousands of dollars and lose all their money,
they're never going to come back. So what is that?
Are those scripts on GitHub? Are they or those
located? GitHub.com slash schematical slash sc
dash terraform. Okay, awesome. I'll give you a
show like it's also I mean, they're all over
schematical.com. So just go there first, I guess.
But yeah, and it's I should note, like this isn't
like it will scale, but this is definitely
designed for the scrappy, smaller startup, you
know, one to five people type thing. So it's
simplified more so than if you had millions of
dollars to throw and you were, you know, had
academics just fallen off the walls, then that
would be a different setup. This one's designed
for the scrappy entrepreneurs. Right. It's not
redundant data center, it's not failovers,
things like that, that a seven digit production
system would have to have. Speaking of tech
stacks, what's the tech stack behind drawn by
drawn by AI? Like what, what code do you use? I'm
sure like where's a hosted things like that? No JS.
We're using the serverless framework. It's
hosted the most the web interactions you get are
hosted, I should like aim in here, are hosted on
Lambda via the API gateway there. So it's actually
this bottom stack down here. We're not using ECS
yet because that's like an always on system. If you
use ECS or it can auto scale, but we're we're saving
a lot of money just doing the lambdas, the bigger
stuff that pumps it into so once you finish
uploading your pictures to s three there, it then
pumps an event into Kinesis. Kinesis then puts
that out to a worker, which then, you know,
consumes that pushes that into an AWS batch queue
and AWS batch is where they manage just one off job.
So if I just just one training job or one I want to
generate this, that boots up if there's not
already one running because it's warm, you know,
because if it's got a whole stack of things, we just
keep the EC two instances on for it and it just eats
through it. That finishes off. And when it's done,
it puts out an event back to Kinesis, which then
sends the email out or fires off a Lambda, which
then sends out the email saying, Hey, great, your
stuff is ready. So that's the long and short of it. I
mean, what made you get into learning about Amazon
products so much like their tech stack in 2010 was
the first I was, I was just a general developer for a
startup and they needed someone to migrate their
tech to AWS because they were on the old school. You
buy a server and you go to a server farm and you put it
in and hope no one trips over it. And I've just been
in it ever since. It's something that excels with
me. I don't, I, I, I can, I live in this stuff. I close
my eyes and I almost touch, you know, Kinesis does
this. I don't know. It's just the way my brain
works. Yeah, that's awesome. I mean, it's a
gigantic time saver. If you can dig in and figure
stuff out like that. So we'll say it's not bad
because it's a skill that's in demand. So it's
another thing. Well, there's so many companies
built on top of like Oz products as well. Like for
example, like all the, you know, back to the email
newsletter, we use Beehive and they send out
emails every day. I'm sure they use, what's the
email system for Aussie and S, S or SQS or something
like that? SES is AWS is one simple email service.
So I'm sure they're using that or some Google
product or something. So there, S, yeah, I see it.
SNS is simple notification service. If you wanted
to send your phone notification to stuff.
Somewhere in here, I've got SES. I think it's down
there, but But there's a tool for everything
pretty much is the whole story. It's an amazing
ecosystem by itself. Yeah. It's kind of a funny
story is that's a double-edged sword because
every time I think it's like, I've got an idea for
something that Amazon hasn't done yet, then I
start building it and then they come out with a
service that does it next week and I'm like, nice.
Schematical, where do you envision like, what are
you looking for the next couple of years down the
road? Are you looking to build Schematical or do a
project like drawn by AI and focus on that? Or
what's your outlook and goal? I would love to build
more scalable products. The consulting
business, I went through a big change this year. I
don't do any more hourly work. And so I mainly
advise and oversee clients and train their teams.
So their teams get trained to do this stuff, not
just me. Otherwise, then there's a limit to how
much I can help. But if I train up other people,
again, rising tides, they can do better. So that's
actually, since I've made that change, it's given
me more time to work on other projects. So I intend
to continue to launch projects like this until
it's not fun anymore. And so yeah, I'm probably
going to consult quite a bit still, but it's no
longer, because I changed the model, I'm not
sitting there like, oh, I got to squeeze every hour
out of this one. It's like, okay, like, what's
what's best for the client? Cool. All right,
everybody's trained. Everybody's working good.
All right, I'm off to the next thing. And they'll
just give you a call if they need help at that point.
Yeah, I do office hours. So there's set times when
they can, but if they need help in between, I answer
the phone. But typically, at least they know when
to come at me with problems. Don't get me started on
consulting pricing and all that stuff. I could do a
whole other podcast on that. I cut it off from
there. Pacific thoughts there. But is this your
full time gig? Pretty much? Oh, yeah. No, I was the
only income I've got is from Schematical.com. And
honestly, I think we're still in the red on draw and
buy right now. So I'm glad the consulting stuff
there. Well, the, I mean, that's one thing that is
the hardest part of anything is like, you have the
greatest pro you have the greatest idea and even
implement it. But if you don't have the marketing
behind it, it's tough. I mean, it's hard to get
customers. I mean, we're, we have, you know, it's
hard to get email subscribers or with any
business. So that's, it's like, that's half the
business right there gaining your customers. But
it will have like, it will grow. It's just got to
stick with it, you know, no, there's been good and
steady traffic. And, you know, every once in a
while, we see a spike like we saw when you guys hit us
up the other day. And I just remember my phone
blowing up going again, who's doing something.
And then I started, you know, looking at the
referral URLs, and I pretty quickly tracked it
down. Well, that's the beautiful thing about the
way we design this thing is even if we don't have
customers for say a month, the system just spins
itself down and doesn't cost me anything. So it's,
we can sit here damn near forever and wait for
customers. So it keeps us in the game longer, you
know, that's awesome. So what's the base cost?
Like if no customers come, what do you have to do?
Pay for like an EC two instance at the bare minimum
or like storage? We're not even paying for an EC two
instance. If they don't put anything in the
lambdas don't invoke the, you know, no, in EC two
instance boots up, all I've got is storage costs.
But if, you know, if you don't come back and sign up,
I'll delete, you know, we can just delete the
models. We don't, we don't need four gigabytes of
models costing, forgive it, I just did the math on
all this too. But yeah, so that's that's it.
There's a little bit for VPC that's always on, but
that's also running my other stuff,
mathematical.com, which is generating paying
the bills. So, you know, it's costs covered by
that. Quite, you know, we asked you a question when
you thought about where AI is headed in the future.
Also, there's all these like doomsday scenarios.
What do you think about AI taking over the world and
ending humanity? Like what are your thoughts on
that? Is that you think that's real? Is that going
to happen eventually? What do you guys think? I
could see something dystopian happening at some
point, depending upon how loose the guardrails
get around AI and decision making. I mean, AI is
going to continue to get smart. Deep fakes are
becoming less distinguishable all the time. Do I
think it's going to happen? Do I know? But could I
imagine scenarios where it does? Yeah, I think so.
Whether it's direct actions, meaning that the AI
itself gets access to whatever it's looking to get
access to, or it finds a way to manipulate people to
get it where it wants to go. And that's where I said
it was mentioning that the deep fakes are coming
out. So, could AI be smart enough to deep fake a
person to encourage something to happen? The
scenario exists, yes. My concerns are less with
like terminator type scenarios and more with the
concept of AI girlfriends or significant others.
And the fact that the rise in popularity there
scares me a little bit for how distance people have
already become over the last few years. I mean,
everybody's used to doing everything over
screens already. And so, you've got the concept of
that your only human interactions, your most
intimate human interactions are with a being that
is completely virtual, things like that. That
scares me a little bit more. I think that could have
a greater effect than we would contemplate. We've
got the survival instinct to not let the computers
march down the street and kick our butts, but
slowly over time, people get lonely and do weird
stuff with that. So, I don't know. It's a little bit
more depressing. Blade Runner? Blade Runner or an
episode of Futurama with Lucy Liu? Yeah, it's
funny because a lot of the developers that we've
sat down with say something similar, that this is
the direction that AI is going, some sort of AI
companion, AI friends. Why hang out with somebody
who is maybe going to stab you in the back, maybe,
isn't going to be a good friend when you can hang out
with a perfect friend, one that you get to create,
one that is interested in you all the time, one that
listens all the time, one that knows everything
about what you're interested in. I mean, there's
something to that that's comforting for people to
know that somebody's always going to be there. And
so, that can be dangerous. I mean, there's also
creations of people who have passed away and now,
or we actually featured a tool one time that was an
ex-boyfriend or ex-girlfriend. You uploaded all
the texts from your previous conversations and
then you get to continue to date them based on an AI
predictive of what they'd say, they'd keep the
names and stuff. So, it's like, hold on a second.
There's something missing from moving on from
relationships or reconciling differences. So,
that could be an issue. I see that as an issue as
well. Yeah. I would like to know the stats for that
product, how many of those users eventually
became, got arrested for some type of domestic
situation. Yeah, they show up to their
ex-girlfriend's house and be like, well, your
text to me that you love me again. I got to come back.
It's an episode of Black Mirror right there. Yeah,
but it is an interesting wild world we're living
in. It has the potential to do so many good things,
but it also has the potential to do bad things, but
it also, I mean, kind of the point, Matt was talking
about, it also has the potential to make us perhaps
all just a bit numb. That perfect friend that never
starts a fight, that never does anything, does
that just make you numb because they're never
going to do anything wrong or unexpected or
particularly caring or things that make human
human. Right? So, it's almost like AI is going to
make things so easy that you're not going to
challenge yourself where you need to be
challenged like a human in anything,
relationships, writing, math, whatever it might
be, you know, but we'll see. It's possible, but I'm
going to, I'll throw out another analogy I use is
that when we first got electricity, I'm sure there
were some guys in cabins out in the woods or just
like the heck with that. Like I can, I'll go out and
chop down a tree to start a fire, you know, but at the
same time, like, indoor plumbing, that's another
example. I don't need indoor plumbing. I'm glad we
have indoor plumbing right now. So, it's one of
those things where it's like, does it make us
weaker perhaps, but does it also give us time for us
up to focus on bigger things? So many of us are out
there just struggling to get by and this could, you
know, enable, give us more free time to create in
ways that we couldn't previously. So, I'm, and I'm
not, you know, I'm going to, I want to go that way
because I did go kind of bleak there for a minute.
So, it's got, it could go either way. I mean, you
think there's going to be, there's going to be like
a, like a, like a split at some point where there's
going to be like sex, almost like the Amish where
they just sort of disavow anything AI and think
it's evil and not touch it. Are you going to know
you're touching it? Are you going to know you're
touching it? That's true. Yeah. You have to like,
you have to basically get off the entire electric
grid to not, to know that you're not touching it.
And, you know, your point, Matt, like, I think
there's a big difference though between
something as, something like electricity that
makes such a massive global lift and is AI going to
be something that's highly utility like, like
electricity or is it going to become something
more like Instagram where it does not perhaps
greatly contribute. So, it's neither one of those
examples is a good example, but one of those is a,
you know, electricity pushes, pushes humanity
ahead. Instagram and social media are arguable to
see what the benefits of society that those tools
are. So, I think it just all depends on the
implementations. I don't want to change the
subject. We want to take a step back here. This is
totally random, but I was, you said, you know,
would it be different sex? I was just curious, is
there Amish people in the matrix and they're
plugged in? They don't know they're plugged in.
Like I kind of want to go ask the authors of the
matrix. I'm like, what, what, what, where does
that stand in this platform? Sorry, totally
random there, but that's all right. It's cool to
talk about this kind of stuff. It's like,
sometimes you got to be high on drugs to talk about
it. But not that I do that, but yeah, we live, I live
by Amish country. Like, I think it's the biggest
Amish population in the entire US. It's like an
hour south of me and Cleveland. And it's
interesting going down there because it's just a
whole different world, even just speaking with
them. They talk different. They act different. I
love the Amish because it's so different from the
world I'm from. But yeah, maybe they're, maybe
they're all just some NPCs from the matrix.
Somehow. Kevin Keely, the founder of, I think it's
Keely, the founder of Wired Magazine had a really
interesting take on the Amish. You know, he says
they're some of the best hackers ever because they
can't have electricity in the home. But what they
end up setting up is pneumatics. There's nothing
against pneumatics. So you'll see these guys with
like pneumatic sewing machines and things like
that. So really interesting if you ever get a
chance to read Kevin Keely's take on that. Okay. Is
that, was that an article you said or a book?
Actually, I think that was a podcast he did with Tim
Ferriss on the four hour, I guess the Tim Ferriss
show probably years ago, but still a fascinating
podcast. Yeah, they are definitely incredibly
ingenuitive like in anything they built our
house, my house and you just see the stuff they do.
They don't follow. It's like their brain isn't
where to follow the rules that society tells them
to follow. So they come up with their own things by
themselves, which is cool to see in action. But
yeah, I'll have to check that article out. It just
explained me to myself somehow. I understand
myself better now. You're Amish, Dommie. I must be
Amish. Absolutely. No, I just played by my rules
and it takes me interesting places. Now's the
point to promoting anything you guys want. Links,
we know drawn by AI, but anything else you want to
put out there, we'll be sure to include them in the
in the comments too. Yeah, if you're looking for
amazing ways to see your dog in an entirely new
light, you know, we've got a little tagline. It's
your dog drawn by AI. So it costs nothing to try it
out. We'd love to see you come by and if you chat into
the little chat box in the bottom right corner,
that ends up on my phone. So you are actually
talking to a person. There is no artificial
intelligence whatsoever in our chat box or our
email responses. And if you need help scaling your
AI solution, you've got a, you know, product that
you're trying to get out there and it's costing an
arm and a leg or you can't figure out how to host it at
scale. Hit me up at Schematical.com. It does not
just need to be stable diffusion. I can run pretty
much anything pie torch in there and I'm sure I
could figure out how to slide in a few other things.
So done some LLM work. But yeah, and there's again,
scripts are available for free. But at
Schematical.com and go to the free section and
have at it. Awesome. One thing I just want to
clarify, Matt, because there's a lot of startups
out there, a lot of people on AI that want to do a
startup, you can help them with their general
backend architecture. That's something that you
do as well. Yeah, I can. If you're having trouble
with that scale, I'd be happy to talk to you as well.
But AI seems so much more hot right now.
Everybody's got an AI tool. But yeah, I'm happy to
talk to you either way. And then subscribe to Ryan
and I's newsletter. We've got weekday updates,
Monday through Friday, we have the top stories of
the day. We have the top AI tools of the day,
including tools like drawn by AI, that's drawn by
dot AI, you get to it. We have the hottest tools, the
hottest stories, some community engagement to we
just included this mystery link, which is a new AI
link that you can click on. It takes you to some cool
thing within AI. Sometimes it's a tool, sometimes
it's a video, sometimes it's something new,
sometimes it's something just off the wall, but we
get the mystery link going on right now. And some
other cool stuff for you.