January 30, 2025 • 51 mins
AI works well in the virtual world. That’s partly because the internet provides so much data to train AI models. But there’s no analogous data set for the physical world – and as a result, AI doesn’t work as well there… yet.
Edward Mehr is the co-founder and CEO of Machina Labs. Edward's problem is this: How can you use AI to turn robots from dumb, inflexible machines into skilled, versatile craftsmen?
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:15):
Pushkin.
Speaker 2 (00:20):
AI works amazingly well. It works terrifyingly well even for
virtual things, for words, for pictures, for videos. This is
true in large part because of the Internet. The Internet
provides this wildly abundant, readily available source of words, pictures,
(00:41):
and videos to train AI models. But there is no analogous,
wildly abundant, readily available data set for the physical world.
There is no gargantuan Internet like repository of data that
describes how things move and bend and break in real
(01:02):
physical space. And as a result, we do not yet
have robust AI for the physical But people are working
on it, and if they succeed, they'll change the way
the world works, not just the world as it appears
on our screens, but the actual physical world, the world
(01:23):
where if you drop something on your foot it hurts.
I'm Jacob Goldstein and this is What's Your Problem, the
show where I talk to people who are trying to
make technological progress. My guest today is Edward Mayer. He's
the co founder and CEO of Machina Labs. Edward's problem
(01:47):
is this, how can you use AI to turn robots
from dumb, inflexible machines into skilled versatile craftsmen. Before he
started Machina Labs, Edward worked in the rocket ship business,
first at SpaceX and then at a company called Relativity Space.
And in the rocket business, the word solw firsthand the
(02:10):
problems of traditional manufacturing. It's the kind of problem he's
now trying to solve with AI and robots. It's a
problem called the rigid factory problem. So I've heard you
use this phrase that's interesting to me, and it's the
rigid factory problem. What's the rigid factory problem?
Speaker 1 (02:32):
That main problem with the factories today is that rigidity,
meaning that if you have to build a physical product,
you pretty much have to build a factory that's designed
for it and built for it. There's a lot of
components that goes into the factory, from machinery all the
way to the tooling that is required to build products
that are specifically designed for the geometry for the material
(02:56):
that you're trying to use. The moment you want to
change that, you have to change your factory, which is
a huge investment. You know, I always give an example
from when I was at SpaceX. You know, you think
of SpaceX as a very innovative and it is you
know on the edge of a hardware space in terms
of innovation in the past twenty four years twenty three
(03:17):
four years that they have existed, they have two rocket families.
There's Starship and there's Falcon, right, because at the moment
you decide on diameter of for example, Falcon nine or
the Falcon family in general, the diameter of that core,
it's very hard to change it. A lot of tooling
and machinery specifically built for that diameter. And that's why
for Starship they had to start from scratch.
Speaker 2 (03:37):
Start from scratch, meaning like not just design, but like
the factory itself, like the Factorily had to build a
whole new factory because they wanted to make a different
sized rocket.
Speaker 1 (03:47):
Yes, different size, different material, All the tooling has to change,
right almost almost, Yeah, you have to basically assume building
from scratch, ground up factory. Why does it need to
be there for us to build this new product?
Speaker 2 (04:00):
I heard you describe was this from your own experience
the sort of era at SpaceX when the fact that
you couldn't make the rocket wider led to all these
kind of difficult things people were trying to do to
be like, how can we do all these things under
this fundamental constraint, Like, can you talk a little bit
(04:20):
about that.
Speaker 1 (04:21):
Yeah, this is a lot of conversation happening in twenty twelve,
twenty thirteen, twenty fourteen time when the diameter of the
Falcon nine could not get any larger, and if you
look at actually different Falcon versions, the height of that
vehicle kept going higher, the diameter could not change. So
it was about where what space can you find to
(04:42):
put new features and new designs that exist within the vehicle.
So there was a lot of stuff basically being crammed
into the space that you have already got.
Speaker 2 (04:53):
So that's true for building rockets. I mean, what are
some other just you know, different kinds of manufactured products
where that kind of rigidity is a problem.
Speaker 1 (05:02):
Yeah, I think it is just common almost in all manufacturing.
That's why this phenomenon. I think it's kind of funny.
People take it for granted that a thing called economies
of scale, uh huh, Like people take it for granted
as if it's rule of nature. It's actually not.
Speaker 2 (05:19):
Just to be clear, it's basically, the more you build
of a thing, the cheaper each one of those things gets.
If you build one, it's really expensive. If you build
a million each one's.
Speaker 1 (05:27):
A lot cheaper, yes, exactly, But then people don't think
about it's like, oh, okay, intuitively makes sense, but why
it's actually not that intuitive. It's actually a limitation of technology. Right.
Why why economies of scale is a thing is because
you have to make a huge amount of investment to
make the first thing, and the moment you make the
second thing and the third thing, then you can break
(05:48):
even your investment onto more products that you're going to
come out of it. But that's only true if the
second product can be built for the first investment. You
had to turn this concept and say, oh, this is
a given. This is an axiom of the world that
economies of scale is a thing, but in reality it
is a technological challenge. Right. It means that you build
a car, you ask what application you Once you build
(06:10):
a factory for a car and all the toolings and
dies that goes into stamping of the panels of that car,
that's one hundred and fifty million dollar investment just for stamping.
And this is a number for example from Tesla. Tesla
spens one hundred and fifty million dollars in a stamping
plant they have in Giga factory in Texas, right, and
that can only make Model Y or Model three right
(06:33):
the moment you have to change that. That means go
through every eighty two hundred and thirty sheet metal panels
that exist on that car and design a new tool
for it. And each of these tool is going to
be a few hundred thousand dollars to sometimes a million
dollars or a million and a half million dollars. And
you're talking about like eighty t one hundred and thirty
tools per vehicle.
Speaker 2 (06:53):
And like all you're doing you're not reinventing the car there,
you're just making a car that's a slightly different shape.
Speaker 1 (06:58):
Basically, yes, maybe you may get sedan, and you're not
trying to do a slightly longer version, a slightly bigger version.
And that's why economies a scale of the thing you saying, Okay,
I made a fact. Now it only pays back if
I make a million of this car, right, because I
had to just drop one hundred and fifty million dollars
on just a stamping plant. So yeah, it's all over manufacturing.
(07:20):
We abstract this whole concept and gave it the name,
says economies, of scale.
Speaker 2 (07:24):
Yeah, so you left SpaceX and you went to Relativity Space, right,
a company that was also in the space business that
was using three D printing.
Speaker 1 (07:34):
Right.
Speaker 2 (07:35):
That was the idea of the company, which seems like
an approach to this problem that you're talking about.
Speaker 1 (07:40):
Right.
Speaker 2 (07:40):
An advantage of three D printing is that it is
much more flexible and less rigid than traditional manufacturing. Right,
So tell me about that.
Speaker 1 (07:50):
Yeah. So, yeah, we saw this challenge at SpaceX and
I joined Relativity very early on. I was the fourth
person on that team. And the goal over there was, Okay,
let's just think about this fundamentally, can we build a
rocket that all built with flexible technology and a time?
Three D printing was that forefront of everybody's minds because
(08:14):
people were already starting to build that. NASA SpaceX people
were already starting to build engines out of three D printing,
and the concept was like, well, that's great, it's very flexible.
Three D printing has this promise of geometry agnostic, material agnostic.
You can just feed it a design and can build
a product for you. And it worked very well with
rocket engines. I think probably the future all rocket engines
(08:36):
would be would be three D printed, And the concept was,
can we take this and scale it to the whole vehicle, right,
can we build the whole vehicle with a process like
three D printing so that it is flexible Today, if
you want to build a rocket with twelve foot diameter,
we can do it. And then if our calculation changes
and we wanted to go to another orbit or do
(08:57):
a different type of emission, then we can change that
diet twelve diameter to twenty diameter twenty fie.
Speaker 2 (09:03):
Don't have to build a new factory, don't have to
build new machinery, just change three D printer.
Speaker 1 (09:07):
Yeah, exactly. So that was a concept behind relativity. That's
a thesis behind relativity, and that was the goal there.
The goal was, you know, three D print a whole
rocket so they can be flexible.
Speaker 2 (09:19):
But it hasn't. It hasn't worked at least in the
kind of maximalist version, right, Like, they just haven't been
able to do it. They've they've sort of backed off
of that that big dream, as I understand it.
Speaker 1 (09:30):
Yeah, yeah, So I think the challenge was that three
D printing is just one process and it's necessarily not
good for every type of part. You know, manufacturing is
very versatile. You do different types of geometries, different types
of material, and three D printing has a very small reach.
(09:50):
There's certain type of parts like rocket engines, very good fit.
You're building a tank, maybe not so right. So yeah
it's good for certain type of parts, but there as a
whole lot of other parts. Like I said, you know
you're building a fuel tank, which is basically large sheet
metal or thin walled structure, then maybe three D printing
is not as good as a fit because it takes
(10:11):
a long time, and also because it's thin, you have
a lot of physical challenges in terms of controlling the
geometry and the tolerances. So we realize soon that maybe
other processes are also need to be automated the same
way three D printing is. We need to have more
flexible processes that are not just one process, more flexible platforms.
(10:34):
They can do different types of processes, not just three
D printing, to be able to cover a whole variety
of products in a flexible manner, the same way the
three D printing that's for certain type of products. And
that was actually the thinking behind MARKETA Labs is that, okay,
can we step back and say, what do we need
to build? What is this flexible platform that can do
(10:57):
three D printing if needed, or it can do sheet
forming if it's needed, It can do machining if it's needed,
but chooses the right operation, right flexible operation for the
right part, but still very agile and doesn't require a
lot of tooling and it's not inflexible.
Speaker 2 (11:12):
So it's it's sort of zooming out more. It's saying
three D printing is not going to do everything the
way manufacturing works now. It's just too rigid, too hard
to change things, to rely on on scale to make
the economics work out. So like that's a very big,
very abstract thought. To start a company, you got to
(11:33):
make something or you got to make something that makes
something like what do you what do you actually do?
Speaker 1 (11:39):
Yeah? So it was interesting, right, you know we actually
the solution was in our past. Right if you look
at like.
Speaker 2 (11:48):
The lesson in a movie, it's like the Wizard of
Oz or something exactly.
Speaker 1 (11:53):
If you look at manufacturing, I mean up to Industrial Revolution,
it was arts and crafts. Right, it was basically humans
trying to figure out how to conquer nature, right, Like,
how am I gonna use my hands? On my brains
and very few primitive tools to deform a product or
shape a product from raw material. Right and to this state,
(12:16):
if you are in a very high mixed manufacturing still
a lot of that creativity exists. There is a person
at Space IX. His name is Big John. I don't
think he's there anymore, but there was this guy. It
was like, you know, a very skilled maker, a craftsman.
You could figure out how to use simpler tools to
build different things in a creative way. Maybe it's not
(12:37):
a repeatable way like you know a stamping works or
ejection molding works, but you can be flexible. You can
do different types of things. You can be creative about
it and do different type of things. So the inspiration
came from how actually humans used to do manufacturing but
realized in order to be flexible, you actually need two components.
You need intelligence and you need set of simple tools
(13:00):
with a lot of kinematic freedom. Now you can pick
up those simple tools, and as long as you have
the intelligence on how to use tools and what sequence
and what kind of a process how to use those tools,
you can actually do a whole variety of projects.
Speaker 2 (13:14):
And so when you say kinematic freedom, you basically mean
like robot arms that can move in lots of different ways.
Is that practically what kinematic freedom means in this context?
Speaker 1 (13:23):
Yes, basically can apply these tools in a lot with
a lot of freedom to the material, right the same
way humans humans do it, right, you know as a human,
you know, if you think about it, you can pick
up a welder and weld something, and then you drop
the welder, and you pick up a drill and you
put a hole in it, and you drop to drill
and you pick up a you know, hammer and maybe
(13:43):
hammer it into shape. So you actually have a few
set of tools, but you have a lot of good
kinematic freedom and most importantly, very creative mind too tells
you how to apply these tools to the material, so
they can actually get very complex set of products and
a lot of diversity.
Speaker 2 (13:59):
So plainly, instead of big John, you want a robot, right,
That's where that's the kinematic freedom. The tools are kind
of like old tools, but optimized for the root. And
then when you say intelligence, that's the one where it's
like feels more frontier ish, like does that mean like
clever engineers figuring out how to automate the robots doesn't
mean AI? Does it mean both?
Speaker 1 (14:21):
Yeah? So I think yeah, you're basically getting to the
crux of how do you scale it? Right? You need
to have those three components, and how does the intelligent piece,
which is the most important piece, comes into play in
an automated fashion. So early days we started from basic
intelligence of humans. But then we had a plan to
capture data and train AI so that you can replace
(14:44):
the thinking and the creativity that human had to put in.
Speaker 2 (14:47):
What's the first thing you decided to try and build.
What's the first sort of problem you want to solve?
Speaker 1 (14:52):
Yeah? I think so. I left relativity in twenty eighteen,
and the idea when I left relativity was there, right.
I was like, okay, we need to build basically what
I had in My mom called it robot craftsmen. Robocraftsman,
we call it a time. How can you build a
robot system? To your point, you can pick up different tools,
has the same king mean, but also have to have
the intelligence. The challenge is you know you said, in
(15:13):
order to train these robots with AI, you need to
have a lot of data. And this is not the
data you can find on internet.
Speaker 2 (15:20):
Right, this is the AI robotics problem, it seems right,
like unlike with large language models, like that's why we
have large language models and not AI robots, right because
because we have the data just sort of randomly sitting
around on the Internet, and we don't have that physical
world data for robots.
Speaker 1 (15:36):
Right exactly. So basically the problem narrowed down into Okay,
how can I generate enough data? How can I create
a business that has a sustaining way of generating data
so I can actually build these models, I can build
this intelligence for these robots. And the thinking was, Okay,
I need to create a solution that can scale in
(15:58):
the industry with limited amount of data and some heuristic.
But then because it's scaling, we can generate a lot
of data and it starts building AI mods.
Speaker 2 (16:09):
Right. You need a first thing that you can actually
do before you really have AI, to generate the data
that will get you to.
Speaker 1 (16:17):
AI exactly exactly. So we're thinking about, Okay, it needs
to be a large enough market right where we can
get mass adoption, and we need to solve a problem
that's big enough it's ten times at least better than
the current solution so it can actually get adoption, right.
Speaker 2 (16:34):
Meaning you can't just do something as well, you have
to do it ten times better.
Speaker 1 (16:38):
Yeah, Because I think what we realize is that through
the last two companies, if something is not ten times better,
cannot overcome the inertia that exists in an industry for
adoption because you know, if you're doing something for the
same way, and in manufacturing, people have been doing things
the old way for hundred of years, right.
Speaker 2 (16:56):
Yeah, and it's a risk, right if they're going to
try working with you, they're immediately taking a risk. And
if it's only going to be a little better, why
should I take that risk?
Speaker 1 (17:05):
Exactly? So the idea was, Okay, we need to find
it large enough market for our first application, and we
need to have a solution that at least ten times better.
So that landed us. We actually looked at a lot
of things, from three D printing to forging to a
lot of things, and then landed on sheet metal. So
sheet metal is the largest metal processing sector out of
all It's a two hundred and eighty billion dollar industry today,
(17:27):
and forming complex sheet metal shapes is very tool intensive.
So so what we started to do was, okay, can
we make our robot craftsman's first operation to be forming
sheet metal, basically forming sheet metal the same way a
sheet shaper hammer is a sheet into shape.
Speaker 2 (17:43):
And when I think about sheet metal, I mean I
don't know anything about sheet metal. I think of like
I think of cars, I think of planes, right, I
think of like you know, detroit, like stamping.
Speaker 1 (17:53):
Is that?
Speaker 2 (17:53):
Am I thinking about the right thing? Am I missing
huge or huge sheet metal universe?
Speaker 1 (17:57):
Like?
Speaker 2 (17:57):
What's the sheet metal universe?
Speaker 1 (17:59):
Yes? So sheep metal almost is everywhere. I think is
the most common metal part that you see on day
to day, right, because most of the time we use
metal to be a container for other things. So it's
usually a thin metal structure that's formed in complex shape
to hold something else. Now you know it can be
from case of a computer. Uh, you know, to a car, right,
(18:23):
you know you're sitting in a freeway you're in to
see a sheet metal or to a to a airplane
you're in a sheet metal can to a rocket body.
Speaker 2 (18:31):
Uh.
Speaker 1 (18:32):
For for a lot of rocket someone will composites with
a lot of a machine metal. And to agricultural heavy
equipment machinery you think of combines, tractors to even building equipments.
You look at your h ract ducts are all sheet metal, right,
because it just makes sense. It's we mostly use metal
(18:53):
parts to contain other things, and we give it complex
shapes and that's where she forming comes into play. So
you pretty much see it everywhere. But the challenge is
that in almost in all cases, you have to create tooling.
It goes back to that first problem. He said, you
have to create tooling for each of those geometries. And
that's why you know a Ford needs to make sure
they can sell a million of an FUN fifty before
(19:15):
they can invest in a plant that makes a new
version of FN fifty, right.
Speaker 2 (19:19):
Because you basically have to build a bespoke factory just
to shape sheet metal in a new way exactly for
a new geometry, for for a new design. Exactly where
is that a particular problem? Like where is it? Where
is that? Where does that acutely bind the fact that
sheet metal is so hard to do if you're not
working at.
Speaker 1 (19:39):
Scale so expensive? Yeah, so I think now you're coming
to the even the third stage of how do you
scale this technology? You need to first find you know,
you said you need to be ten ex better we
need Right, you're in an area that has a lot
of pain with today's time.
Speaker 2 (19:54):
I was like, oh my god, thank god, you've walked
through the door. We've been waiting for you.
Speaker 1 (19:58):
Yeah, So end up being very much defense in airspace. Right,
So think of you know, think of our military for example,
right today, they we have fifty sixty different weapon system
or defense systems you can basically think of like aircrafts
that they're maintaining. And some of these systems have been
(20:19):
built from sixty seventy eighty years ago, like think of
B fifty two C one thirty like World War.
Speaker 2 (20:25):
Two planes still flying kind.
Speaker 1 (20:27):
Of still flying, yes, exactly, and they have like you know,
thirty of one, fifty of another, one hundred of another one.
And these things get break down, right, and unlike a
Ford factory, there is no factory for seventy different products
that they're carrying.
Speaker 2 (20:42):
Right, and presumably the factory they built in nineteen forty
one to build this plane doesn't.
Speaker 1 (20:47):
Exist any It doesn't exact. Even the vendor might completely
have disappeared, right, that made that misspecific component. So they're
constantly battling with this challenge of an aircraft goes down,
how can I fix it? How can I find the part?
And there are thousands of parts in each of these aircrafts, right,
so any of them can go down, and that's a
huge challenge. I mean, if you look at you know,
government of a government accountabilit The office put this report out,
(21:11):
I think it was a couple of years ago or
a year ago about how ready each weapon system is
to defend the United States. Out of the forty eight
to forty nine weapon systems they look into only one,
only one in the past eleven years, every year was ready, right.
I think only top four had like at least half
(21:32):
of the years ready right. So that means in most
years these weapons are not ready to fight, like.
Speaker 2 (21:37):
They're waiting for parts.
Speaker 1 (21:39):
They're waiting for parts. Something is broken, something is damaged,
and we used to go deeper. Some of these components
take four years to be replaced. So if a plane
gets damaged, it needs to sit on the ground for
four years before it can be it can be replaced,
and the cost of replacement is building another factory basically,
So some of these parts, and think of it, a
landing gear door that goes on a plane will cost
(22:00):
them eight hundred thousand dollars for example, because they have.
Speaker 2 (22:02):
To go make it because it's bespoke essentially, like buying
a bespoke suit or something. It's just like it's gonna
cost a lot.
Speaker 1 (22:08):
Yeah, yeah, So the idea started there. I think that
was one of our first customers. Can we make defense
manufacturing more agile? Directly affects our national readiness for military
conflict and it's a huge problem. But then you know,
even in a broader sense, any defense product or aerospace
product usually has very low volume but high mix of products.
(22:29):
You know, even you know, you're building a missile, you
make like, you know, a few thousand a year, and
you might make five, six seven different versions of right,
So it's very unlike cars, where you know, you make
a million of the same car over and over all.
So that ends up being our first application, which we've
got a lot of traction with. But but you know,
even outside of that, you know, you look at companies
(22:50):
like Caterpillar, like John Deere's of the world. These folks
also are in the same book. You know, they make
two hundred combines, right, but they need to support them
in the field. And these folks have the exactly same problem, right,
you know, do I need to run a large factory
to support all these models at all time, and that's
will be very expensive to support, like one hundred vehicles out.
Speaker 2 (23:11):
There still to come on the show. We'll talk about
the future of AI and robotics at Mocking the Labs
and beyond. And so you got the right market. Now
(23:33):
you've got to make a thing. You got to figure
out how to actually do the thing, how to make
your idea come true? Like how does that work?
Speaker 1 (23:41):
So the idea originally was can we get rid of
a die right and do it the same way a
sheet shaper forms a sheet of metal? And what does
a sheet chaper do? Sheet chaper get starts from a
flat sheet of metal and it slowly hammers it into shape.
So what we wanted to do was have a robot
do that, right, have robotic system basically do that incremental
(24:03):
deformation into shape. We call it romophor.
Speaker 2 (24:05):
So you're sort of bending it, right, I mean you're
hammering it. It's sort of like if you take a whatever,
cut open a luminum can and kind of bend it
into shape. Like that's a version of what's happening here, right, Yes,
exactly complicated way, yeah.
Speaker 1 (24:17):
Exactly, you're right, I mean the same way a potter
forms a clay bowl. That's basically what our robots do.
They start from a flash sheet of metal and slowly
deforming in the shape the same way a potter, which
form it clay bowl or a sheet shape of hammets
and hammers a sheet into shape. Yeah, so I've seen
it right.
Speaker 2 (24:34):
So there will be a sheet of metal like hanging
hanging up in whatever above the ground. And then you
have a robot arm on either side, right, like one
on one side, one on the other. And then what.
Speaker 1 (24:44):
Happens Basically the robots come together from both sides the
sheet and they pinch the sheet in a certain way
so that that location that they're pinching slightly stretches and deforms. Right.
And if you start applying this pinching all over the
sheet and incrementally, you slowly start to form it into
(25:05):
a shape. Right. So instead of traditionally would use it
die and with sheer pressure of the press pushing the
sheet against the dye to give it a shape. Now
the robots are like a craftsman, like a trades person
coming in to slowly deform the sheet into shape by
just applying pressure. So one robot is pushing it the
other robot is supporting it, and by applying a pinch
(25:26):
you slightly stretch the material and you form it into
a shape.
Speaker 2 (25:30):
So I mean, the way you describe it, it makes
sense and it sounds easy. I'm sure it wasn't easy,
Like were there things that just didn't work for a while.
Speaker 1 (25:44):
So you should have been here when the first time
we actually tried to form a part, the part looked
like it was like a ghost of the geometry that
they wanted to make, and actually in the end it
tore right. So think about it. You have this very
flimsy sheet applying pressure to it, and if features apply
pressure slightly wrong right, it can potentially tear it. It
(26:08):
can form it into a different shape. And also the
whole sheet is moving the whole time you're trying to
move to form it. The whole sheet is moving because
it's very flimsy. It's not a rigid structure, right. So
the main challenge was how do you get this accurate? Right?
How do you get this process accurate? How do you
get accuracy? And the idea was what does the robot
need to do given all of these chaotic nature of
(26:30):
the process where the sheet moves and if you apply
it too much pressure, it will deform in a bad
way or in my tear. If you're probably not enough pressure,
it might just not form. So how do you come
up with the right set of robot movements and process
parameters to form the part? And that was the problem
we want to solve with AI right, but we didn't
have the data right right in the beginning. Right. The
(26:52):
idea was that if I form enough parts with this process,
and I can capture all the data throughout the process,
where did the robot go, how much pressure did it apply,
and what was the resulting geometry? That can start building
a model that says that correlates the inputs to the outputs,
and I can explore this and say, okay, in order
to get to the right output, I need these inputs.
(27:12):
But we didn't have them in the beginning. So the
idea was two things. One was maybe we can simulate
the data right and very early on we started doing
some simulation, physics based simulation, and we soon realized in
order to get an accurate result, the simulations are going
to be very computationally intensive. A simulation of a part
that took only fifteen minutes to form took us one
(27:35):
week on twenty seven core machine. Wow. Right, so okay,
stimulation not only is not accurate, it takes forever. So
we realized, okay, so that's not the right route. The
right route was like, Okay, we can also form a
lot of parts and gather the data. But in order
to do that we go back to that same problem.
We need to have a scale. We need to have
a lot of these machines for these parts and get
(27:56):
that data.
Speaker 2 (27:57):
I mean, one of the big AI insights of the
last whatever decade is like, you need a ton of data,
which is easy if it's words, but hard if it's metal. Right.
Speaker 1 (28:07):
Yes, we ended up doing was created a hybrid model.
We said, okay, what if we keep the humans in
the loop, so the human can give an instruction initially
based on herostricts, and then we look at the data
and human can adjust and then iterate on that. But
while we are capturing all these data, and over time,
(28:29):
as we're capturing the data, we start building the models
that will help the human do less trials. Right. It's
basically guided reinforcement learning, right, and a humans are actually
guiding it where to go, but it's exploring those areas
but after a while, once we started forming south thousands
of parts, then you can start feeding this data into model.
Then the model will be like, okay, human, you don't
(28:50):
need to do twenty five different trials. Now you can
do with five trials, you're going to get to the
right place, which is actually the number we are at
right now.
Speaker 2 (28:56):
And that's happening in the physical world largely those iterations
like you're trying a piece of metal and it's bad
and it tears, and you do another piece of metal
and it's a little less bad and eventually.
Speaker 1 (29:06):
Exactly exactly, and that initially would take twenty five parts,
like you know, before we find a recipe for that design.
But twenty five parts still was better than traditional alternative.
Speaker 2 (29:16):
When you say twenty five parts, I mean twenty five
tries twenty five pieces of metal before you make the
part the right way exactly.
Speaker 1 (29:23):
And that was like, you know, they would sit down
basically twenty five days in a row, so in a
month they could actually define a recipe where traditionally making
a mold would take at least three four months. Right,
So we were still better. But then now with over time,
when we generated the data and now the model can
tell the engineer, okay, maybe you want to choose these parameters,
is now becoming an advisor with down to five trials.
(29:46):
In five trials, we can actually get to the right
part and then hopefully in the future we get to
a point where you know, the machine will tell the
romans what to do and the human can be completely
out of the loop. Yeah. But the idea was like,
how do you kind of create that hybrid model that's
efficient so that we can generate the data until the
model is good enough to do the job itself.
Speaker 2 (30:05):
And you find that the data is sort of generalizable,
I mean clearly, like making one kind of part makes
the model the AI smarter about making another kind of part.
Speaker 1 (30:16):
Yes, you know, yeah it is. It's kind of interesting.
I think people don't think about it. I used to
do sheet shaping by hand, right, That was one of
the hobbies I had. I was working with this shop
in Pomona that we were actually hammer sheets into shape,
and we used to say, you know, if you spent
five years doing it, you're really good. You get really
good at it. I was used to think, you know, okay,
after five years of doing this, yes, you have this
(30:38):
intuitive understanding of you look at the sheet and be like, okay,
this this place needs to be hammered more. This place
needs to be hammered. It was, it was, it was.
It was intuitive. It was like you couldn't explains why
you're thinking this need to happen. There was no physical explanation.
None of these people who were she shaping got PhDs
in material science. Yeah, they just learned over time seeing
the pattern of how the sheet formed. Yes, craftsmanship, that's
(30:58):
craftsmanship right. Yeah, but really reminded me of Okay, these
people can know how to do it, but without really
being able to explain it, to do it for five years.
Speaker 2 (31:07):
It's that kind of tacit knowledge.
Speaker 1 (31:09):
Yeah, and reminded me of the same challenge we had
early machine learning challenge where they were like, okay, a
human can look at two pictures and say, okay, this
is a cat and this is a dog. Something happens
in their brain that knows which is a cat, but
they cannot really define why they're calling this cat and
this sort of dog. So that was where it starts
to click for me. If I can capture enough data,
(31:30):
five years worth of data right of a human, then
I should be able to get to a very good
sheet shaper, right, And you know it's funny. Back at
the end, I was like, okay, humans are you know,
receiving x amount of megabytes a second? Okay, how five
years worse of data? Is that much? So roughly, I
think once we get a certain amount of data, I
think we have enough data to be able to basically
(31:53):
replace a like not replace replace the mentality or the
model that the sheet shaper has in their mind.
Speaker 2 (31:59):
So how how how many years of kind of human
level craftsmen sheet shaping data does the model have at
this point?
Speaker 1 (32:07):
Yeah? No, so I think rastam I check? One year ago?
I checked around, we were like three fourths of the
way there in terms of the data that we have
for just she shaping. Right. So once we get to
I think full, I think and these at that point
we have no excuse. We have enough data. The model
should be good. We just need to figure out how
why it's not. Maybe far from it is.
Speaker 2 (32:24):
It is interesting to analogize it to like human craftsmanship, right,
And I mean even if you want to zoom out
even more, the like fifty year history of AI, where
first everybody was like, oh, you just got to teach
the machine all the rules for to use your example,
like what's a cat and what's a dog? But then
you realize it's actually wildly hard to make a list
of rules that can reliably distinguish a cat from a dog.
(32:48):
And the weird thing that has happened in AI is like, oh,
you don't actually have to make a list. You just
need like image that you just need like a giant
database of images and a giant neural network and you
just throw it at it like and say figure it out,
and it figures it out, and you're sort of doing that.
(33:09):
But for shaping metal.
Speaker 1 (33:11):
For metal, and then the only challenge was, like you know,
cats and dogs pictures were Internet and sheet metal forming
data wasn't. And so that's that was an additional problem
we have to solve, as you pointed out, which is
a big problem in physical AI.
Speaker 2 (33:23):
So I want to talk a little bit more about
AI and robotics. Jansen Wong has been talking about it,
as I'm sure you know in video and videos vc
ARM as an investor in your company, other people are
working on what you're working on. I mean, I'm curious
what does the sort of AI and robotics path look
like to you? For the next few years, and what like,
(33:45):
what do you understand about it now that you didn't
understand whatever five years ago? Like what what have you
really come to realize by working on it all the time?
Speaker 1 (33:54):
I think the biggest problem for physical AI is data
generation for now to train models. So we need to
either there's two things need to happen. Either new types
of models needs to be created, new architectures, new new
algorithms basically, which I'm sure it's going to happen that
can learn more with less data basically and the same
(34:15):
way humans kind of learn more with less data. Right.
But at the same time, I think, you know, we
only exposed our models to categorically to ten percent of
type of data that humans receive. You know, you think
about you know, human intrictions. You and I are now talking,
if it was AI, AI is probably only listening to
the words we're saying, right, But that's only ten percent
(34:38):
of communication. I can see your lips moving, I can
see your eyebrows moving. I can see like maybe you're
folding your arms and okay, I know that like okay,
maybe there's all these ninety percent of the signals are
not captured. That that's used for learning. You know, you
look at if you ask chat GPT or Dolly or
you know any of the you know, even even you
know Grock say okay, draw me a clock that is
(35:02):
shows five thirty. It cannot show you draw you a clock.
It will draw you a clock, but it doesn't show
five thirty. Actually, most the time it shows ten ten.
Speaker 2 (35:10):
Ten ten, because that's where watchhands, like analog watchhands look good.
Speaker 1 (35:14):
Right, it's a nice little v because those are all
the images that they're seen on internet because they watch it.
Speaker 2 (35:19):
It's almost always ten ten. It's the classic watch photo.
Speaker 1 (35:22):
It's like five thirty is also ten ten, because.
Speaker 2 (35:24):
It's always ten ten right to a generative AI, it's
always ten ten somewhere.
Speaker 1 (35:30):
So I think, but that humans, you know, receive this
data of movement. When you grow up you look at
the clock on the wall as a kid, you're like, okay,
now I intuitively get it. I think I know what's
going on, so I can actually make it work. So
even though we train it a lot of data, I
don't think we trained it on the right categorically right
data yet right to get all the intuitive understanding that
(35:50):
we have today. So I think we have a data
problem and that exists the physical AI. So I think
the applications will win. There's a lot of people are
working in this. I think the applications will win who
can either synthetically generate that data or they can actually
scale in the physical world in a way where they
can actually generate the day for themselves. But the scaling
(36:12):
needs to happen with less data, and I think that
was That's why I'm like, for example, like very bullish
on manufacturing. So I think the data is going to
be the biggest challenge. And I think, you know, in
order for us to massively change this space, we need
to be able to get to the data. I don't
think algorithms is a bottleneck there yet. It's just a
data for us.
Speaker 2 (36:31):
And is it just a matter of people doing what
you're doing and like finding little wedge places to start
and having people sort of hold the hand of the
model and training up the models. I mean that seems
slow on a certain level, like not you know, obviously
it's working for you, but like, is there some kind
(36:52):
of breakthrough move people can make? Can you put sensors
somewhere in the world to you know, train AI without
having to you know, have a human stand next to
it as it messes up one piece of sheet metal
after another.
Speaker 1 (37:05):
Yeah, I think I think that there is there's another path,
which is simulation path. Make physics based simulations faster and
kind of learn. Let the robots just go play in
a digital playground as opposed to deploy it in real role,
and that becomes a computation problem. And then you know,
as long as you have enough computation, you can train
to robots. But I think, you know, I think you
(37:26):
know the good examples that we have had such success
so far as like autonomous cars, right, did the same
thing we were doing, but in the car like Okay, Tesla,
you know, deploy the fleet of robots that are capturing
data still be driven by humans, but the data can
be used later on to kind of automate it.
Speaker 2 (37:42):
I mean, that's an interesting case because it has been
much harder clearly than many people thought. Maybe most people thought, right, Like,
I know, that's a particular instance where you're really worried
about edge cases. I don't know, is autonomous cars like
a good model or not. It seems complicated.
Speaker 1 (38:01):
I think the model of capturing data is there, but
then the the task at hand is very hard. Yeah, right,
so I think that's the challenge, right so where it
says like with us, it's still much more structured environment,
And I think that's that was the thinking we're thinking.
I think the hardest problem right now in physical AI
is finding the business model of how do you scale
data capture without requiring billions of dollars in investment?
Speaker 2 (38:24):
So what do you make in today?
Speaker 1 (38:26):
I imagine you know, so last time I checked in
the facility, one four of the sales are working on
a defense application.
Speaker 2 (38:35):
Is it secret? Can you tell me what it is?
Speaker 1 (38:38):
It's a missile? And two of them were working on
an aerospace application. This is components of an aircraft or
a drawing. And one of them, as an interesting one,
was working on an architectural component, which is a roof
tile for a specific building that's used by the Department
of the by Bureau of Water Recognition.
Speaker 2 (38:58):
Oh, I was I was going to say, what is it?
Something like Frank Gary, like nightmare weirdo metal park.
Speaker 1 (39:05):
Oh, those we have had those in the past two
but this one is actually very practical. Well, it's this building.
It's actually very interesting. Exactly these buildings, these large industrial
buildings that built they built in the sixties or fifties,
and they use these type of roof tiles that the
manufacturer doesn't exist anymore. And anybody else who they went
to quoted them hundreds of thousand dollars to make those tiles,
(39:26):
and we're like, oh no, we can make it for you.
But also that show is kind of the diversity. I mean,
like like I say, in the morning, we have like
aerospace parts. In the afternoon, roof tiles for a industrial complex,
for you know, for a dam.
Speaker 2 (39:40):
Now you're in the sheet metal business. I know you're large.
Dream is much larger than that, right, but like what
like that, tell me where you are now? Tell me
where you are now? Like what are you doing? What
are you selling? And then kind of what's the next
big step.
Speaker 1 (39:57):
So some of our systems are now operating out in
the wild and working for the customers. And but I
think the next phase of growth for us is getting
into each of these applications and own more of the
process so we can teach the robocraftsmen the future processes
not just sheet for me, but also maybe how to
assemble it, how to weld it, how do you surface
(40:17):
finish it right. So what we are doing now in
the next phase is actually instead of selling parts or
components or systems, we're actually saying, Okay, can we get
this robocraftsman to actually build you a subassembly or a
full product, not just a component of it, but a
full product. So that's something we're describing with folks. Can
(40:39):
we have the robocraftsmen build the full drone for you?
Can we have the robocrafts and build you a full
missile as opposed to just build missile you know missile scans.
Speaker 2 (40:49):
Is there that seems like a leap? Is there not
an intermediate step?
Speaker 1 (40:55):
Like yes? Yes? So I mean how we're doing it
is we're gradually stepping into it right the same way
she metal was our first application. So we're putting a
facility that maybe makes drones, but the main component that
we automate today is sheet for me, which is the bottleneck.
And then we do the welding in a traditional way
on the same robots, but we actually instruct them to
(41:17):
do it.
Speaker 2 (41:18):
So that way, the robot is kind of back where
it was on sheet metal five years ago, but it's
learning how to weld now exactly.
Speaker 1 (41:26):
I used to work in a you know, a shop
that we will do custom cars, build the custom cars
with hand, and so it was also near near and
dear to my heart. So what we realize is that
with our technology, for the first time, we can actually
enable a product that didn't exist in automotive, meaning that
instead of buying a car that's mass produced and every
(41:48):
single one of them look the same, you can now
let the customer design a custom car for them. You know,
right now, if you go buy a car, you can
you have options of what the what the seat color
would be, or maybe the color of the car would be,
and what some trim options. But you can't really choose
the design of your car. You can't say, oh, I
want a different hood and I want a different fender
(42:09):
because going to the back same problem you have to
make tooling and mold for the vender of certain designs.
It cannot easily change it. So with our technology you can.
So what we started doing was like, okay, applying this
freedom that this technology provides to now automotive is the
ability of the customers to be able to go to
a website design a fully customized car for themselves. It
(42:33):
can be either from already design panels round car designer
or adding a specific customer customizations they want to do,
for example, logo of their company to their door of
the car or the hood of the car, and actually
get a completely unique car right manufactured for them. And
we're actually working with this with some of our automotive
partners Automotive Aims as well. Right, we actually showed some
(42:55):
of this work in the biggest aftermarket show in the
United States is called SEMA with our partner Toyota. So
I think this is going to be, in my opinion,
one of the new product categories in automotive. We have
had a time of his cars, we have had you know,
electric cars, and I think now for the first time,
with technologies like ours, you can have custom to order cars,
(43:17):
like cars that are like, you know, the same way
you choose what T shirt you wear and your T
shirt is different than mine. We also don't have to
drive the same you know model S or you know
Model three. We can actually have our own customized Model
three is and modelesses.
Speaker 2 (43:30):
So what's the I mean is that the if you
think sort of long term for Makeina is like that
what you think about like give me the give me
the five year vision yeah, or ten year or whatever.
Speaker 1 (43:43):
Yeah. So I think the long term motivation behind our
company is can you grant this democratization of ideas for
people who want to build anything? Right? Can I express
myself if I'm a builder, can I go build something
with not having to build a factory for So that's
really the long term goal. So I imagine in the
next five to ten years, you can as a designer,
(44:07):
somebody who has an idea, you can go to a website,
get guided through your ideas on how to make and
design a physical product, hit a button and say, okay,
I want twenty of these, and I want in Chatsworth, California,
and the right facility programs the right number of robots
to actually do those operations without any hardware or investment
(44:29):
that needs to be made for those of specific parts
and ship it to you two days later in the
right location. That is the future we're building towards cars
is just you know, one of the products that could
be built. But I imagine that you know this technology
or technology like these, technologies like these can be used
to do the myriad of designs. I think the moment
you open up this possibility of any designs could be
(44:52):
a reality. I think so many things will be created
that we're not even thinking of right now. You know
the fact that we have cars today and they all
look the same as limitation of technology. But the moment
you can open up this creativity of turning ideas into
physical reality without a without an initial investment or huge
barrier to entry, then I think we're going to have
(45:13):
all kinds of drones, all kinds of satellites, all kinds
of rockets, all kinds of cars that you're going to
be this like you know, Cambrian explosion of different designs
that's going to come into our world. And I think
that's what future is about. The future is about you
know what I call it, Like future is custom Like
future is about being able to make these all these
ideas in reality. We had this explosion happening in digital world. Yeah,
(45:33):
you know, now we have even models generating images and
videos and there's this you know, explosion of different ideas
and content being created using the technology. But the link
is broken to the physical world and the physical work
is still pretty uniform because it's very hard to make
things in the physical world. Can we bridge that gap?
Can we connect the digital world of creation to physical
(45:56):
world of creation and create the same variety in the
physical world as we have in the digital world. I
think that's the goal in our company.
Speaker 2 (46:06):
We'll be back in a minute with the light Year Round.
Let's finish with the Lightning Round. Do you drive a
customized car?
Speaker 1 (46:24):
I don't actually yet. Well, if I am, what have I.
Speaker 2 (46:27):
Seen on your Instagram? What's that truck you keep posting
on your Instagram?
Speaker 1 (46:31):
So I so I have a truck that's customized. I
don't drive it around as much, but maybe this year
I'll start taking it out. This year I've been you know,
we have been kind of stelf about it, talking about it,
but we haven't talked about it in a big way
because we have a big release coming soon.
Speaker 2 (46:45):
I mean, you're literally posting it on Instagram. It's not
that stuff. Tell me about Tell me about that truck
you keep posting on Instagram? What's going on with that?
Speaker 1 (46:53):
So it's a truck is fully the full body is
fully customized.
Speaker 2 (46:56):
It says anvil in the back when you post it?
Is it called anvil?
Speaker 1 (46:59):
Dumb question and call it and call it anvil? I
think it's the idea was actually the shape design of
it was inspired by amil. If you look at the
front fender, it actually looks like a the front bumper
looks like an anvil. But also the idea is that, like,
you know, we're actually forming shits on an anvil. So yeah,
it was very fitting.
Speaker 2 (47:15):
Tell me about that truck, like, tell just tell me
what's it look like.
Speaker 1 (47:18):
Yeah, so for example, like you know, we put a
lot of form and sharp edges in the in the hood. Right.
Most vehicles have a very hard time if you look
at most of the you know, hood of the vehicles,
they are very smooth because it's very hard to actually
put sharp angles in the hood. So if you look
at this truck, this truck has a lot of angles,
a lot of sharp detail right in the hood. Right,
(47:41):
And and and that's very expressive of the type of
person for example, that I am, right, I like things
that are edgy, and and that truck is certainly edgy. Right.
It's bare metal, right, you know, there is no blemishes
being hiden and hidden under the under the vehicle. You know,
a lot of people when cyber Truck came out, we
(48:02):
got very excited about you know, oh, it's bare metal.
It looks like a metal, but then there was no
form in it because it's actually very hard to make
it for metal look nice. And so that's one of
the things we wanted to show. We want to show that, Okay,
you can actually have a form metal with a lot
of detail in it and still keep it bare metal
because it will look nice. Right, So, yeah, a lot
(48:22):
of design features of it. For me kind of represents
the type of personality and character that I have. But
I think that's how every car should be. You know,
people should be able to have that freedom to choose
what their cars look like.
Speaker 2 (48:34):
How many skull tattoos do you have?
Speaker 1 (48:38):
I've got three? Why it's it's so, yeah, it's an
interesting thing. So a skull for me represents kind of
and it's an abstract for death of ego. So I
have a tattoo on my thumb which is a skull
that's holding a microphone to his ears. And this was
(49:01):
a time where you know, I felt like, you know,
I had a good platform and I could talk a
lot and people would listen. But then I realized I
should yes, that's right, but I should maybe keep them
keep the mic close to my ears and also listen
as opposed to talk all the time, right, So I think.
Speaker 2 (49:17):
Skull microphones don't work that way for the record, but.
Speaker 1 (49:21):
I like it as a metaphor exactly. But I think
the idea of is around, you know, kind of reminders
of you can see a lot of my tattoos on
my on my hands, so it's a really reminder for
myself to know that, you know, be present and make
sure that you know you're not involved with your ego
too much and you can see others people's perspective.
Speaker 2 (49:41):
Is there any tension between ego death and custom cars?
Speaker 1 (49:46):
Tension between ego death and custom cars? I don't know.
Speaker 2 (49:50):
I'm just playing, but like you know, custom car kind
of seems like, hey, look at me, I'm special, and
ego death seems like, oh, don't look at me, I'm
not so special.
Speaker 1 (49:59):
Yeah, no, I think I think the difference is I think, yeah,
if you have attachment to your custom card, then maybe
there's tension. But I more think of it in terms
of expression. Right. You know, you can be an artist,
you can. You can you can design your home the
way it expresses you. You can design the theme of
your podcast the way it expresses you. You can design
your car. Also, the way it expresses you. I think
(50:21):
it's leus so about oh look at me, I'm special.
It's more like, here's my expression to the world for
the people to see. But I think that expressiveness is
it is pretty amazing. I think that's uniquely one of
the unique things about humans that like, you know, we
we I think all we do when we come to
this world is expressing ourselves right, expressing uself through our work,
expressing through ourselves through our relationships. And if you can
(50:45):
enable people to express themselves better better, I think that's great.
But if you get attached to your expressions and your
ideas and your thoughts and think, oh, I'm better than
everybody else, and I think that that becomes that becomes
a little bit of an ego driven trip.
Speaker 2 (51:05):
Edward Mayer is the co founder and CEO of Mocking Labs.
Today's show was produced by Gabriel Hunter Cheang. It was
edited by Lyddy jeene Kott and engineered by Sarah Bruginner.
You can email us at problem at Pushkin dot FM.
I'm Jacob Goldstein and we'll be back next week with
another episode of What's Your Problem.
Pushkin.
Speaker 2 (00:20):
AI works amazingly well. It works terrifyingly well even for
virtual things, for words, for pictures, for videos. This is
true in large part because of the Internet. The Internet
provides this wildly abundant, readily available source of words, pictures,
(00:41):
and videos to train AI models. But there is no analogous,
wildly abundant, readily available data set for the physical world.
There is no gargantuan Internet like repository of data that
describes how things move and bend and break in real
(01:02):
physical space. And as a result, we do not yet
have robust AI for the physical But people are working
on it, and if they succeed, they'll change the way
the world works, not just the world as it appears
on our screens, but the actual physical world, the world
(01:23):
where if you drop something on your foot it hurts.
I'm Jacob Goldstein and this is What's Your Problem, the
show where I talk to people who are trying to
make technological progress. My guest today is Edward Mayer. He's
the co founder and CEO of Machina Labs. Edward's problem
(01:47):
is this, how can you use AI to turn robots
from dumb, inflexible machines into skilled versatile craftsmen. Before he
started Machina Labs, Edward worked in the rocket ship business,
first at SpaceX and then at a company called Relativity Space.
And in the rocket business, the word solw firsthand the
(02:10):
problems of traditional manufacturing. It's the kind of problem he's
now trying to solve with AI and robots. It's a
problem called the rigid factory problem. So I've heard you
use this phrase that's interesting to me, and it's the
rigid factory problem. What's the rigid factory problem?
Speaker 1 (02:32):
That main problem with the factories today is that rigidity,
meaning that if you have to build a physical product,
you pretty much have to build a factory that's designed
for it and built for it. There's a lot of
components that goes into the factory, from machinery all the
way to the tooling that is required to build products
that are specifically designed for the geometry for the material
(02:56):
that you're trying to use. The moment you want to
change that, you have to change your factory, which is
a huge investment. You know, I always give an example
from when I was at SpaceX. You know, you think
of SpaceX as a very innovative and it is you
know on the edge of a hardware space in terms
of innovation in the past twenty four years twenty three
(03:17):
four years that they have existed, they have two rocket families.
There's Starship and there's Falcon, right, because at the moment
you decide on diameter of for example, Falcon nine or
the Falcon family in general, the diameter of that core,
it's very hard to change it. A lot of tooling
and machinery specifically built for that diameter. And that's why
for Starship they had to start from scratch.
Speaker 2 (03:37):
Start from scratch, meaning like not just design, but like
the factory itself, like the Factorily had to build a
whole new factory because they wanted to make a different
sized rocket.
Speaker 1 (03:47):
Yes, different size, different material, All the tooling has to change,
right almost almost, Yeah, you have to basically assume building
from scratch, ground up factory. Why does it need to
be there for us to build this new product?
Speaker 2 (04:00):
I heard you describe was this from your own experience
the sort of era at SpaceX when the fact that
you couldn't make the rocket wider led to all these
kind of difficult things people were trying to do to
be like, how can we do all these things under
this fundamental constraint, Like, can you talk a little bit
(04:20):
about that.
Speaker 1 (04:21):
Yeah, this is a lot of conversation happening in twenty twelve,
twenty thirteen, twenty fourteen time when the diameter of the
Falcon nine could not get any larger, and if you
look at actually different Falcon versions, the height of that
vehicle kept going higher, the diameter could not change. So
it was about where what space can you find to
(04:42):
put new features and new designs that exist within the vehicle.
So there was a lot of stuff basically being crammed
into the space that you have already got.
Speaker 2 (04:53):
So that's true for building rockets. I mean, what are
some other just you know, different kinds of manufactured products
where that kind of rigidity is a problem.
Speaker 1 (05:02):
Yeah, I think it is just common almost in all manufacturing.
That's why this phenomenon. I think it's kind of funny.
People take it for granted that a thing called economies
of scale, uh huh, Like people take it for granted
as if it's rule of nature. It's actually not.
Speaker 2 (05:19):
Just to be clear, it's basically, the more you build
of a thing, the cheaper each one of those things gets.
If you build one, it's really expensive. If you build
a million each one's.
Speaker 1 (05:27):
A lot cheaper, yes, exactly, But then people don't think
about it's like, oh, okay, intuitively makes sense, but why
it's actually not that intuitive. It's actually a limitation of technology. Right.
Why why economies of scale is a thing is because
you have to make a huge amount of investment to
make the first thing, and the moment you make the
second thing and the third thing, then you can break
(05:48):
even your investment onto more products that you're going to
come out of it. But that's only true if the
second product can be built for the first investment. You
had to turn this concept and say, oh, this is
a given. This is an axiom of the world that
economies of scale is a thing, but in reality it
is a technological challenge. Right. It means that you build
a car, you ask what application you Once you build
(06:10):
a factory for a car and all the toolings and
dies that goes into stamping of the panels of that car,
that's one hundred and fifty million dollar investment just for stamping.
And this is a number for example from Tesla. Tesla
spens one hundred and fifty million dollars in a stamping
plant they have in Giga factory in Texas, right, and
that can only make Model Y or Model three right
(06:33):
the moment you have to change that. That means go
through every eighty two hundred and thirty sheet metal panels
that exist on that car and design a new tool
for it. And each of these tool is going to
be a few hundred thousand dollars to sometimes a million
dollars or a million and a half million dollars. And
you're talking about like eighty t one hundred and thirty
tools per vehicle.
Speaker 2 (06:53):
And like all you're doing you're not reinventing the car there,
you're just making a car that's a slightly different shape.
Speaker 1 (06:58):
Basically, yes, maybe you may get sedan, and you're not
trying to do a slightly longer version, a slightly bigger version.
And that's why economies a scale of the thing you saying, Okay,
I made a fact. Now it only pays back if
I make a million of this car, right, because I
had to just drop one hundred and fifty million dollars
on just a stamping plant. So yeah, it's all over manufacturing.
(07:20):
We abstract this whole concept and gave it the name,
says economies, of scale.
Speaker 2 (07:24):
Yeah, so you left SpaceX and you went to Relativity Space, right,
a company that was also in the space business that
was using three D printing.
Speaker 1 (07:34):
Right.
Speaker 2 (07:35):
That was the idea of the company, which seems like
an approach to this problem that you're talking about.
Speaker 1 (07:40):
Right.
Speaker 2 (07:40):
An advantage of three D printing is that it is
much more flexible and less rigid than traditional manufacturing. Right,
So tell me about that.
Speaker 1 (07:50):
Yeah. So, yeah, we saw this challenge at SpaceX and
I joined Relativity very early on. I was the fourth
person on that team. And the goal over there was, Okay,
let's just think about this fundamentally, can we build a
rocket that all built with flexible technology and a time?
Three D printing was that forefront of everybody's minds because
(08:14):
people were already starting to build that. NASA SpaceX people
were already starting to build engines out of three D printing,
and the concept was like, well, that's great, it's very flexible.
Three D printing has this promise of geometry agnostic, material agnostic.
You can just feed it a design and can build
a product for you. And it worked very well with
rocket engines. I think probably the future all rocket engines
(08:36):
would be would be three D printed, And the concept was,
can we take this and scale it to the whole vehicle, right,
can we build the whole vehicle with a process like
three D printing so that it is flexible Today, if
you want to build a rocket with twelve foot diameter,
we can do it. And then if our calculation changes
and we wanted to go to another orbit or do
(08:57):
a different type of emission, then we can change that
diet twelve diameter to twenty diameter twenty fie.
Speaker 2 (09:03):
Don't have to build a new factory, don't have to
build new machinery, just change three D printer.
Speaker 1 (09:07):
Yeah, exactly. So that was a concept behind relativity. That's
a thesis behind relativity, and that was the goal there.
The goal was, you know, three D print a whole
rocket so they can be flexible.
Speaker 2 (09:19):
But it hasn't. It hasn't worked at least in the
kind of maximalist version, right, Like, they just haven't been
able to do it. They've they've sort of backed off
of that that big dream, as I understand it.
Speaker 1 (09:30):
Yeah, yeah, So I think the challenge was that three
D printing is just one process and it's necessarily not
good for every type of part. You know, manufacturing is
very versatile. You do different types of geometries, different types
of material, and three D printing has a very small reach.
(09:50):
There's certain type of parts like rocket engines, very good fit.
You're building a tank, maybe not so right. So yeah
it's good for certain type of parts, but there as a
whole lot of other parts. Like I said, you know
you're building a fuel tank, which is basically large sheet
metal or thin walled structure, then maybe three D printing
is not as good as a fit because it takes
(10:11):
a long time, and also because it's thin, you have
a lot of physical challenges in terms of controlling the
geometry and the tolerances. So we realize soon that maybe
other processes are also need to be automated the same
way three D printing is. We need to have more
flexible processes that are not just one process, more flexible platforms.
(10:34):
They can do different types of processes, not just three
D printing, to be able to cover a whole variety
of products in a flexible manner, the same way the
three D printing that's for certain type of products. And
that was actually the thinking behind MARKETA Labs is that, okay,
can we step back and say, what do we need
to build? What is this flexible platform that can do
(10:57):
three D printing if needed, or it can do sheet
forming if it's needed, It can do machining if it's needed,
but chooses the right operation, right flexible operation for the
right part, but still very agile and doesn't require a
lot of tooling and it's not inflexible.
Speaker 2 (11:12):
So it's it's sort of zooming out more. It's saying
three D printing is not going to do everything the
way manufacturing works now. It's just too rigid, too hard
to change things, to rely on on scale to make
the economics work out. So like that's a very big,
very abstract thought. To start a company, you got to
(11:33):
make something or you got to make something that makes
something like what do you what do you actually do?
Speaker 1 (11:39):
Yeah? So it was interesting, right, you know we actually
the solution was in our past. Right if you look
at like.
Speaker 2 (11:48):
The lesson in a movie, it's like the Wizard of
Oz or something exactly.
Speaker 1 (11:53):
If you look at manufacturing, I mean up to Industrial Revolution,
it was arts and crafts. Right, it was basically humans
trying to figure out how to conquer nature, right, Like,
how am I gonna use my hands? On my brains
and very few primitive tools to deform a product or
shape a product from raw material. Right and to this state,
(12:16):
if you are in a very high mixed manufacturing still
a lot of that creativity exists. There is a person
at Space IX. His name is Big John. I don't
think he's there anymore, but there was this guy. It
was like, you know, a very skilled maker, a craftsman.
You could figure out how to use simpler tools to
build different things in a creative way. Maybe it's not
(12:37):
a repeatable way like you know a stamping works or
ejection molding works, but you can be flexible. You can
do different types of things. You can be creative about
it and do different type of things. So the inspiration
came from how actually humans used to do manufacturing but
realized in order to be flexible, you actually need two components.
You need intelligence and you need set of simple tools
(13:00):
with a lot of kinematic freedom. Now you can pick
up those simple tools, and as long as you have
the intelligence on how to use tools and what sequence
and what kind of a process how to use those tools,
you can actually do a whole variety of projects.
Speaker 2 (13:14):
And so when you say kinematic freedom, you basically mean
like robot arms that can move in lots of different ways.
Is that practically what kinematic freedom means in this context?
Speaker 1 (13:23):
Yes, basically can apply these tools in a lot with
a lot of freedom to the material, right the same
way humans humans do it, right, you know as a human,
you know, if you think about it, you can pick
up a welder and weld something, and then you drop
the welder, and you pick up a drill and you
put a hole in it, and you drop to drill
and you pick up a you know, hammer and maybe
(13:43):
hammer it into shape. So you actually have a few
set of tools, but you have a lot of good
kinematic freedom and most importantly, very creative mind too tells
you how to apply these tools to the material, so
they can actually get very complex set of products and
a lot of diversity.
Speaker 2 (13:59):
So plainly, instead of big John, you want a robot, right,
That's where that's the kinematic freedom. The tools are kind
of like old tools, but optimized for the root. And
then when you say intelligence, that's the one where it's
like feels more frontier ish, like does that mean like
clever engineers figuring out how to automate the robots doesn't
mean AI? Does it mean both?
Speaker 1 (14:21):
Yeah? So I think yeah, you're basically getting to the
crux of how do you scale it? Right? You need
to have those three components, and how does the intelligent piece,
which is the most important piece, comes into play in
an automated fashion. So early days we started from basic
intelligence of humans. But then we had a plan to
capture data and train AI so that you can replace
(14:44):
the thinking and the creativity that human had to put in.
Speaker 2 (14:47):
What's the first thing you decided to try and build.
What's the first sort of problem you want to solve?
Speaker 1 (14:52):
Yeah? I think so. I left relativity in twenty eighteen,
and the idea when I left relativity was there, right.
I was like, okay, we need to build basically what
I had in My mom called it robot craftsmen. Robocraftsman,
we call it a time. How can you build a
robot system? To your point, you can pick up different tools,
has the same king mean, but also have to have
the intelligence. The challenge is you know you said, in
(15:13):
order to train these robots with AI, you need to
have a lot of data. And this is not the
data you can find on internet.
Speaker 2 (15:20):
Right, this is the AI robotics problem, it seems right,
like unlike with large language models, like that's why we
have large language models and not AI robots, right because
because we have the data just sort of randomly sitting
around on the Internet, and we don't have that physical
world data for robots.
Speaker 1 (15:36):
Right exactly. So basically the problem narrowed down into Okay,
how can I generate enough data? How can I create
a business that has a sustaining way of generating data
so I can actually build these models, I can build
this intelligence for these robots. And the thinking was, Okay,
I need to create a solution that can scale in
(15:58):
the industry with limited amount of data and some heuristic.
But then because it's scaling, we can generate a lot
of data and it starts building AI mods.
Speaker 2 (16:09):
Right. You need a first thing that you can actually
do before you really have AI, to generate the data
that will get you to.
Speaker 1 (16:17):
AI exactly exactly. So we're thinking about, Okay, it needs
to be a large enough market right where we can
get mass adoption, and we need to solve a problem
that's big enough it's ten times at least better than
the current solution so it can actually get adoption, right.
Speaker 2 (16:34):
Meaning you can't just do something as well, you have
to do it ten times better.
Speaker 1 (16:38):
Yeah, Because I think what we realize is that through
the last two companies, if something is not ten times better,
cannot overcome the inertia that exists in an industry for
adoption because you know, if you're doing something for the
same way, and in manufacturing, people have been doing things
the old way for hundred of years, right.
Speaker 2 (16:56):
Yeah, and it's a risk, right if they're going to
try working with you, they're immediately taking a risk. And
if it's only going to be a little better, why
should I take that risk?
Speaker 1 (17:05):
Exactly? So the idea was, Okay, we need to find
it large enough market for our first application, and we
need to have a solution that at least ten times better.
So that landed us. We actually looked at a lot
of things, from three D printing to forging to a
lot of things, and then landed on sheet metal. So
sheet metal is the largest metal processing sector out of
all It's a two hundred and eighty billion dollar industry today,
(17:27):
and forming complex sheet metal shapes is very tool intensive.
So so what we started to do was, okay, can
we make our robot craftsman's first operation to be forming
sheet metal, basically forming sheet metal the same way a
sheet shaper hammer is a sheet into shape.
Speaker 2 (17:43):
And when I think about sheet metal, I mean I
don't know anything about sheet metal. I think of like
I think of cars, I think of planes, right, I
think of like you know, detroit, like stamping.
Speaker 1 (17:53):
Is that?
Speaker 2 (17:53):
Am I thinking about the right thing? Am I missing
huge or huge sheet metal universe?
Speaker 1 (17:57):
Like?
Speaker 2 (17:57):
What's the sheet metal universe?
Speaker 1 (17:59):
Yes? So sheep metal almost is everywhere. I think is
the most common metal part that you see on day
to day, right, because most of the time we use
metal to be a container for other things. So it's
usually a thin metal structure that's formed in complex shape
to hold something else. Now you know it can be
from case of a computer. Uh, you know, to a car, right,
(18:23):
you know you're sitting in a freeway you're in to
see a sheet metal or to a to a airplane
you're in a sheet metal can to a rocket body.
Speaker 2 (18:31):
Uh.
Speaker 1 (18:32):
For for a lot of rocket someone will composites with
a lot of a machine metal. And to agricultural heavy
equipment machinery you think of combines, tractors to even building equipments.
You look at your h ract ducts are all sheet metal, right,
because it just makes sense. It's we mostly use metal
(18:53):
parts to contain other things, and we give it complex
shapes and that's where she forming comes into play. So
you pretty much see it everywhere. But the challenge is
that in almost in all cases, you have to create tooling.
It goes back to that first problem. He said, you
have to create tooling for each of those geometries. And
that's why you know a Ford needs to make sure
they can sell a million of an FUN fifty before
(19:15):
they can invest in a plant that makes a new
version of FN fifty, right.
Speaker 2 (19:19):
Because you basically have to build a bespoke factory just
to shape sheet metal in a new way exactly for
a new geometry, for for a new design. Exactly where
is that a particular problem? Like where is it? Where
is that? Where does that acutely bind the fact that
sheet metal is so hard to do if you're not
working at.
Speaker 1 (19:39):
Scale so expensive? Yeah, so I think now you're coming
to the even the third stage of how do you
scale this technology? You need to first find you know,
you said you need to be ten ex better we
need Right, you're in an area that has a lot
of pain with today's time.
Speaker 2 (19:54):
I was like, oh my god, thank god, you've walked
through the door. We've been waiting for you.
Speaker 1 (19:58):
Yeah, So end up being very much defense in airspace. Right,
So think of you know, think of our military for example,
right today, they we have fifty sixty different weapon system
or defense systems you can basically think of like aircrafts
that they're maintaining. And some of these systems have been
(20:19):
built from sixty seventy eighty years ago, like think of
B fifty two C one thirty like World War.
Speaker 2 (20:25):
Two planes still flying kind.
Speaker 1 (20:27):
Of still flying, yes, exactly, and they have like you know,
thirty of one, fifty of another, one hundred of another one.
And these things get break down, right, and unlike a
Ford factory, there is no factory for seventy different products
that they're carrying.
Speaker 2 (20:42):
Right, and presumably the factory they built in nineteen forty
one to build this plane doesn't.
Speaker 1 (20:47):
Exist any It doesn't exact. Even the vendor might completely
have disappeared, right, that made that misspecific component. So they're
constantly battling with this challenge of an aircraft goes down,
how can I fix it? How can I find the part?
And there are thousands of parts in each of these aircrafts, right,
so any of them can go down, and that's a
huge challenge. I mean, if you look at you know,
government of a government accountabilit The office put this report out,
(21:11):
I think it was a couple of years ago or
a year ago about how ready each weapon system is
to defend the United States. Out of the forty eight
to forty nine weapon systems they look into only one,
only one in the past eleven years, every year was ready, right.
I think only top four had like at least half
(21:32):
of the years ready right. So that means in most
years these weapons are not ready to fight, like.
Speaker 2 (21:37):
They're waiting for parts.
Speaker 1 (21:39):
They're waiting for parts. Something is broken, something is damaged,
and we used to go deeper. Some of these components
take four years to be replaced. So if a plane
gets damaged, it needs to sit on the ground for
four years before it can be it can be replaced,
and the cost of replacement is building another factory basically,
So some of these parts, and think of it, a
landing gear door that goes on a plane will cost
(22:00):
them eight hundred thousand dollars for example, because they have.
Speaker 2 (22:02):
To go make it because it's bespoke essentially, like buying
a bespoke suit or something. It's just like it's gonna
cost a lot.
Speaker 1 (22:08):
Yeah, yeah, So the idea started there. I think that
was one of our first customers. Can we make defense
manufacturing more agile? Directly affects our national readiness for military
conflict and it's a huge problem. But then you know,
even in a broader sense, any defense product or aerospace
product usually has very low volume but high mix of products.
(22:29):
You know, even you know, you're building a missile, you
make like, you know, a few thousand a year, and
you might make five, six seven different versions of right,
So it's very unlike cars, where you know, you make
a million of the same car over and over all.
So that ends up being our first application, which we've
got a lot of traction with. But but you know,
even outside of that, you know, you look at companies
(22:50):
like Caterpillar, like John Deere's of the world. These folks
also are in the same book. You know, they make
two hundred combines, right, but they need to support them
in the field. And these folks have the exactly same problem, right,
you know, do I need to run a large factory
to support all these models at all time, and that's
will be very expensive to support, like one hundred vehicles out.
Speaker 2 (23:11):
There still to come on the show. We'll talk about
the future of AI and robotics at Mocking the Labs
and beyond. And so you got the right market. Now
(23:33):
you've got to make a thing. You got to figure
out how to actually do the thing, how to make
your idea come true? Like how does that work?
Speaker 1 (23:41):
So the idea originally was can we get rid of
a die right and do it the same way a
sheet shaper forms a sheet of metal? And what does
a sheet chaper do? Sheet chaper get starts from a
flat sheet of metal and it slowly hammers it into shape.
So what we wanted to do was have a robot
do that, right, have robotic system basically do that incremental
(24:03):
deformation into shape. We call it romophor.
Speaker 2 (24:05):
So you're sort of bending it, right, I mean you're
hammering it. It's sort of like if you take a whatever,
cut open a luminum can and kind of bend it
into shape. Like that's a version of what's happening here, right, Yes,
exactly complicated way, yeah.
Speaker 1 (24:17):
Exactly, you're right, I mean the same way a potter
forms a clay bowl. That's basically what our robots do.
They start from a flash sheet of metal and slowly
deforming in the shape the same way a potter, which
form it clay bowl or a sheet shape of hammets
and hammers a sheet into shape. Yeah, so I've seen
it right.
Speaker 2 (24:34):
So there will be a sheet of metal like hanging
hanging up in whatever above the ground. And then you
have a robot arm on either side, right, like one
on one side, one on the other. And then what.
Speaker 1 (24:44):
Happens Basically the robots come together from both sides the
sheet and they pinch the sheet in a certain way
so that that location that they're pinching slightly stretches and deforms. Right.
And if you start applying this pinching all over the
sheet and incrementally, you slowly start to form it into
(25:05):
a shape. Right. So instead of traditionally would use it
die and with sheer pressure of the press pushing the
sheet against the dye to give it a shape. Now
the robots are like a craftsman, like a trades person
coming in to slowly deform the sheet into shape by
just applying pressure. So one robot is pushing it the
other robot is supporting it, and by applying a pinch
(25:26):
you slightly stretch the material and you form it into
a shape.
Speaker 2 (25:30):
So I mean, the way you describe it, it makes
sense and it sounds easy. I'm sure it wasn't easy,
Like were there things that just didn't work for a while.
Speaker 1 (25:44):
So you should have been here when the first time
we actually tried to form a part, the part looked
like it was like a ghost of the geometry that
they wanted to make, and actually in the end it
tore right. So think about it. You have this very
flimsy sheet applying pressure to it, and if features apply
pressure slightly wrong right, it can potentially tear it. It
(26:08):
can form it into a different shape. And also the
whole sheet is moving the whole time you're trying to
move to form it. The whole sheet is moving because
it's very flimsy. It's not a rigid structure, right. So
the main challenge was how do you get this accurate? Right?
How do you get this process accurate? How do you
get accuracy? And the idea was what does the robot
need to do given all of these chaotic nature of
(26:30):
the process where the sheet moves and if you apply
it too much pressure, it will deform in a bad
way or in my tear. If you're probably not enough pressure,
it might just not form. So how do you come
up with the right set of robot movements and process
parameters to form the part? And that was the problem
we want to solve with AI right, but we didn't
have the data right right in the beginning. Right. The
(26:52):
idea was that if I form enough parts with this process,
and I can capture all the data throughout the process,
where did the robot go, how much pressure did it apply,
and what was the resulting geometry? That can start building
a model that says that correlates the inputs to the outputs,
and I can explore this and say, okay, in order
to get to the right output, I need these inputs.
(27:12):
But we didn't have them in the beginning. So the
idea was two things. One was maybe we can simulate
the data right and very early on we started doing
some simulation, physics based simulation, and we soon realized in
order to get an accurate result, the simulations are going
to be very computationally intensive. A simulation of a part
that took only fifteen minutes to form took us one
(27:35):
week on twenty seven core machine. Wow. Right, so okay,
stimulation not only is not accurate, it takes forever. So
we realized, okay, so that's not the right route. The
right route was like, Okay, we can also form a
lot of parts and gather the data. But in order
to do that we go back to that same problem.
We need to have a scale. We need to have
a lot of these machines for these parts and get
(27:56):
that data.
Speaker 2 (27:57):
I mean, one of the big AI insights of the
last whatever decade is like, you need a ton of data,
which is easy if it's words, but hard if it's metal. Right.
Speaker 1 (28:07):
Yes, we ended up doing was created a hybrid model.
We said, okay, what if we keep the humans in
the loop, so the human can give an instruction initially
based on herostricts, and then we look at the data
and human can adjust and then iterate on that. But
while we are capturing all these data, and over time,
(28:29):
as we're capturing the data, we start building the models
that will help the human do less trials. Right. It's
basically guided reinforcement learning, right, and a humans are actually
guiding it where to go, but it's exploring those areas
but after a while, once we started forming south thousands
of parts, then you can start feeding this data into model.
Then the model will be like, okay, human, you don't
(28:50):
need to do twenty five different trials. Now you can
do with five trials, you're going to get to the
right place, which is actually the number we are at
right now.
Speaker 2 (28:56):
And that's happening in the physical world largely those iterations
like you're trying a piece of metal and it's bad
and it tears, and you do another piece of metal
and it's a little less bad and eventually.
Speaker 1 (29:06):
Exactly exactly, and that initially would take twenty five parts,
like you know, before we find a recipe for that design.
But twenty five parts still was better than traditional alternative.
Speaker 2 (29:16):
When you say twenty five parts, I mean twenty five
tries twenty five pieces of metal before you make the
part the right way exactly.
Speaker 1 (29:23):
And that was like, you know, they would sit down
basically twenty five days in a row, so in a
month they could actually define a recipe where traditionally making
a mold would take at least three four months. Right,
So we were still better. But then now with over time,
when we generated the data and now the model can
tell the engineer, okay, maybe you want to choose these parameters,
is now becoming an advisor with down to five trials.
(29:46):
In five trials, we can actually get to the right
part and then hopefully in the future we get to
a point where you know, the machine will tell the
romans what to do and the human can be completely
out of the loop. Yeah. But the idea was like,
how do you kind of create that hybrid model that's
efficient so that we can generate the data until the
model is good enough to do the job itself.
Speaker 2 (30:05):
And you find that the data is sort of generalizable,
I mean clearly, like making one kind of part makes
the model the AI smarter about making another kind of part.
Speaker 1 (30:16):
Yes, you know, yeah it is. It's kind of interesting.
I think people don't think about it. I used to
do sheet shaping by hand, right, That was one of
the hobbies I had. I was working with this shop
in Pomona that we were actually hammer sheets into shape,
and we used to say, you know, if you spent
five years doing it, you're really good. You get really
good at it. I was used to think, you know, okay,
after five years of doing this, yes, you have this
(30:38):
intuitive understanding of you look at the sheet and be like, okay,
this this place needs to be hammered more. This place
needs to be hammered. It was, it was, it was.
It was intuitive. It was like you couldn't explains why
you're thinking this need to happen. There was no physical explanation.
None of these people who were she shaping got PhDs
in material science. Yeah, they just learned over time seeing
the pattern of how the sheet formed. Yes, craftsmanship, that's
(30:58):
craftsmanship right. Yeah, but really reminded me of Okay, these
people can know how to do it, but without really
being able to explain it, to do it for five years.
Speaker 2 (31:07):
It's that kind of tacit knowledge.
Speaker 1 (31:09):
Yeah, and reminded me of the same challenge we had
early machine learning challenge where they were like, okay, a
human can look at two pictures and say, okay, this
is a cat and this is a dog. Something happens
in their brain that knows which is a cat, but
they cannot really define why they're calling this cat and
this sort of dog. So that was where it starts
to click for me. If I can capture enough data,
(31:30):
five years worth of data right of a human, then
I should be able to get to a very good
sheet shaper, right, And you know it's funny. Back at
the end, I was like, okay, humans are you know,
receiving x amount of megabytes a second? Okay, how five
years worse of data? Is that much? So roughly, I
think once we get a certain amount of data, I
think we have enough data to be able to basically
(31:53):
replace a like not replace replace the mentality or the
model that the sheet shaper has in their mind.
Speaker 2 (31:59):
So how how how many years of kind of human
level craftsmen sheet shaping data does the model have at
this point?
Speaker 1 (32:07):
Yeah? No, so I think rastam I check? One year ago?
I checked around, we were like three fourths of the
way there in terms of the data that we have
for just she shaping. Right. So once we get to
I think full, I think and these at that point
we have no excuse. We have enough data. The model
should be good. We just need to figure out how
why it's not. Maybe far from it is.
Speaker 2 (32:24):
It is interesting to analogize it to like human craftsmanship, right,
And I mean even if you want to zoom out
even more, the like fifty year history of AI, where
first everybody was like, oh, you just got to teach
the machine all the rules for to use your example,
like what's a cat and what's a dog? But then
you realize it's actually wildly hard to make a list
of rules that can reliably distinguish a cat from a dog.
(32:48):
And the weird thing that has happened in AI is like, oh,
you don't actually have to make a list. You just
need like image that you just need like a giant
database of images and a giant neural network and you
just throw it at it like and say figure it out,
and it figures it out, and you're sort of doing that.
(33:09):
But for shaping metal.
Speaker 1 (33:11):
For metal, and then the only challenge was, like you know,
cats and dogs pictures were Internet and sheet metal forming
data wasn't. And so that's that was an additional problem
we have to solve, as you pointed out, which is
a big problem in physical AI.
Speaker 2 (33:23):
So I want to talk a little bit more about
AI and robotics. Jansen Wong has been talking about it,
as I'm sure you know in video and videos vc
ARM as an investor in your company, other people are
working on what you're working on. I mean, I'm curious
what does the sort of AI and robotics path look
like to you? For the next few years, and what like,
(33:45):
what do you understand about it now that you didn't
understand whatever five years ago? Like what what have you
really come to realize by working on it all the time?
Speaker 1 (33:54):
I think the biggest problem for physical AI is data
generation for now to train models. So we need to
either there's two things need to happen. Either new types
of models needs to be created, new architectures, new new
algorithms basically, which I'm sure it's going to happen that
can learn more with less data basically and the same
(34:15):
way humans kind of learn more with less data. Right.
But at the same time, I think, you know, we
only exposed our models to categorically to ten percent of
type of data that humans receive. You know, you think
about you know, human intrictions. You and I are now talking,
if it was AI, AI is probably only listening to
the words we're saying, right, But that's only ten percent
(34:38):
of communication. I can see your lips moving, I can
see your eyebrows moving. I can see like maybe you're
folding your arms and okay, I know that like okay,
maybe there's all these ninety percent of the signals are
not captured. That that's used for learning. You know, you
look at if you ask chat GPT or Dolly or
you know any of the you know, even even you
know Grock say okay, draw me a clock that is
(35:02):
shows five thirty. It cannot show you draw you a clock.
It will draw you a clock, but it doesn't show
five thirty. Actually, most the time it shows ten ten.
Speaker 2 (35:10):
Ten ten, because that's where watchhands, like analog watchhands look good.
Speaker 1 (35:14):
Right, it's a nice little v because those are all
the images that they're seen on internet because they watch it.
Speaker 2 (35:19):
It's almost always ten ten. It's the classic watch photo.
Speaker 1 (35:22):
It's like five thirty is also ten ten, because.
Speaker 2 (35:24):
It's always ten ten right to a generative AI, it's
always ten ten somewhere.
Speaker 1 (35:30):
So I think, but that humans, you know, receive this
data of movement. When you grow up you look at
the clock on the wall as a kid, you're like, okay,
now I intuitively get it. I think I know what's
going on, so I can actually make it work. So
even though we train it a lot of data, I
don't think we trained it on the right categorically right
data yet right to get all the intuitive understanding that
(35:50):
we have today. So I think we have a data
problem and that exists the physical AI. So I think
the applications will win. There's a lot of people are
working in this. I think the applications will win who
can either synthetically generate that data or they can actually
scale in the physical world in a way where they
can actually generate the day for themselves. But the scaling
(36:12):
needs to happen with less data, and I think that
was That's why I'm like, for example, like very bullish
on manufacturing. So I think the data is going to
be the biggest challenge. And I think, you know, in
order for us to massively change this space, we need
to be able to get to the data. I don't
think algorithms is a bottleneck there yet. It's just a
data for us.
Speaker 2 (36:31):
And is it just a matter of people doing what
you're doing and like finding little wedge places to start
and having people sort of hold the hand of the
model and training up the models. I mean that seems
slow on a certain level, like not you know, obviously
it's working for you, but like, is there some kind
(36:52):
of breakthrough move people can make? Can you put sensors
somewhere in the world to you know, train AI without
having to you know, have a human stand next to
it as it messes up one piece of sheet metal
after another.
Speaker 1 (37:05):
Yeah, I think I think that there is there's another path,
which is simulation path. Make physics based simulations faster and
kind of learn. Let the robots just go play in
a digital playground as opposed to deploy it in real role,
and that becomes a computation problem. And then you know,
as long as you have enough computation, you can train
to robots. But I think, you know, I think you
(37:26):
know the good examples that we have had such success
so far as like autonomous cars, right, did the same
thing we were doing, but in the car like Okay, Tesla,
you know, deploy the fleet of robots that are capturing
data still be driven by humans, but the data can
be used later on to kind of automate it.
Speaker 2 (37:42):
I mean, that's an interesting case because it has been
much harder clearly than many people thought. Maybe most people thought, right, Like,
I know, that's a particular instance where you're really worried
about edge cases. I don't know, is autonomous cars like
a good model or not. It seems complicated.
Speaker 1 (38:01):
I think the model of capturing data is there, but
then the the task at hand is very hard. Yeah, right,
so I think that's the challenge, right so where it
says like with us, it's still much more structured environment,
And I think that's that was the thinking we're thinking.
I think the hardest problem right now in physical AI
is finding the business model of how do you scale
data capture without requiring billions of dollars in investment?
Speaker 2 (38:24):
So what do you make in today?
Speaker 1 (38:26):
I imagine you know, so last time I checked in
the facility, one four of the sales are working on
a defense application.
Speaker 2 (38:35):
Is it secret? Can you tell me what it is?
Speaker 1 (38:38):
It's a missile? And two of them were working on
an aerospace application. This is components of an aircraft or
a drawing. And one of them, as an interesting one,
was working on an architectural component, which is a roof
tile for a specific building that's used by the Department
of the by Bureau of Water Recognition.
Speaker 2 (38:58):
Oh, I was I was going to say, what is it?
Something like Frank Gary, like nightmare weirdo metal park.
Speaker 1 (39:05):
Oh, those we have had those in the past two
but this one is actually very practical. Well, it's this building.
It's actually very interesting. Exactly these buildings, these large industrial
buildings that built they built in the sixties or fifties,
and they use these type of roof tiles that the
manufacturer doesn't exist anymore. And anybody else who they went
to quoted them hundreds of thousand dollars to make those tiles,
(39:26):
and we're like, oh no, we can make it for you.
But also that show is kind of the diversity. I mean,
like like I say, in the morning, we have like
aerospace parts. In the afternoon, roof tiles for a industrial complex,
for you know, for a dam.
Speaker 2 (39:40):
Now you're in the sheet metal business. I know you're large.
Dream is much larger than that, right, but like what
like that, tell me where you are now? Tell me
where you are now? Like what are you doing? What
are you selling? And then kind of what's the next
big step.
Speaker 1 (39:57):
So some of our systems are now operating out in
the wild and working for the customers. And but I
think the next phase of growth for us is getting
into each of these applications and own more of the
process so we can teach the robocraftsmen the future processes
not just sheet for me, but also maybe how to
assemble it, how to weld it, how do you surface
(40:17):
finish it right. So what we are doing now in
the next phase is actually instead of selling parts or
components or systems, we're actually saying, Okay, can we get
this robocraftsman to actually build you a subassembly or a
full product, not just a component of it, but a
full product. So that's something we're describing with folks. Can
(40:39):
we have the robocraftsmen build the full drone for you?
Can we have the robocrafts and build you a full
missile as opposed to just build missile you know missile scans.
Speaker 2 (40:49):
Is there that seems like a leap? Is there not
an intermediate step?
Speaker 1 (40:55):
Like yes? Yes? So I mean how we're doing it
is we're gradually stepping into it right the same way
she metal was our first application. So we're putting a
facility that maybe makes drones, but the main component that
we automate today is sheet for me, which is the bottleneck.
And then we do the welding in a traditional way
on the same robots, but we actually instruct them to
(41:17):
do it.
Speaker 2 (41:18):
So that way, the robot is kind of back where
it was on sheet metal five years ago, but it's
learning how to weld now exactly.
Speaker 1 (41:26):
I used to work in a you know, a shop
that we will do custom cars, build the custom cars
with hand, and so it was also near near and
dear to my heart. So what we realize is that
with our technology, for the first time, we can actually
enable a product that didn't exist in automotive, meaning that
instead of buying a car that's mass produced and every
(41:48):
single one of them look the same, you can now
let the customer design a custom car for them. You know,
right now, if you go buy a car, you can
you have options of what the what the seat color
would be, or maybe the color of the car would be,
and what some trim options. But you can't really choose
the design of your car. You can't say, oh, I
want a different hood and I want a different fender
(42:09):
because going to the back same problem you have to
make tooling and mold for the vender of certain designs.
It cannot easily change it. So with our technology you can.
So what we started doing was like, okay, applying this
freedom that this technology provides to now automotive is the
ability of the customers to be able to go to
a website design a fully customized car for themselves. It
(42:33):
can be either from already design panels round car designer
or adding a specific customer customizations they want to do,
for example, logo of their company to their door of
the car or the hood of the car, and actually
get a completely unique car right manufactured for them. And
we're actually working with this with some of our automotive
partners Automotive Aims as well. Right, we actually showed some
(42:55):
of this work in the biggest aftermarket show in the
United States is called SEMA with our partner Toyota. So
I think this is going to be, in my opinion,
one of the new product categories in automotive. We have
had a time of his cars, we have had you know,
electric cars, and I think now for the first time,
with technologies like ours, you can have custom to order cars,
(43:17):
like cars that are like, you know, the same way
you choose what T shirt you wear and your T
shirt is different than mine. We also don't have to
drive the same you know model S or you know
Model three. We can actually have our own customized Model
three is and modelesses.
Speaker 2 (43:30):
So what's the I mean is that the if you
think sort of long term for Makeina is like that
what you think about like give me the give me
the five year vision yeah, or ten year or whatever.
Speaker 1 (43:43):
Yeah. So I think the long term motivation behind our
company is can you grant this democratization of ideas for
people who want to build anything? Right? Can I express
myself if I'm a builder, can I go build something
with not having to build a factory for So that's
really the long term goal. So I imagine in the
next five to ten years, you can as a designer,
(44:07):
somebody who has an idea, you can go to a website,
get guided through your ideas on how to make and
design a physical product, hit a button and say, okay,
I want twenty of these, and I want in Chatsworth, California,
and the right facility programs the right number of robots
to actually do those operations without any hardware or investment
(44:29):
that needs to be made for those of specific parts
and ship it to you two days later in the
right location. That is the future we're building towards cars
is just you know, one of the products that could
be built. But I imagine that you know this technology
or technology like these, technologies like these can be used
to do the myriad of designs. I think the moment
you open up this possibility of any designs could be
(44:52):
a reality. I think so many things will be created
that we're not even thinking of right now. You know
the fact that we have cars today and they all
look the same as limitation of technology. But the moment
you can open up this creativity of turning ideas into
physical reality without a without an initial investment or huge
barrier to entry, then I think we're going to have
(45:13):
all kinds of drones, all kinds of satellites, all kinds
of rockets, all kinds of cars that you're going to
be this like you know, Cambrian explosion of different designs
that's going to come into our world. And I think
that's what future is about. The future is about you
know what I call it, Like future is custom Like
future is about being able to make these all these
ideas in reality. We had this explosion happening in digital world. Yeah,
(45:33):
you know, now we have even models generating images and
videos and there's this you know, explosion of different ideas
and content being created using the technology. But the link
is broken to the physical world and the physical work
is still pretty uniform because it's very hard to make
things in the physical world. Can we bridge that gap?
Can we connect the digital world of creation to physical
(45:56):
world of creation and create the same variety in the
physical world as we have in the digital world. I
think that's the goal in our company.
Speaker 2 (46:06):
We'll be back in a minute with the light Year Round.
Let's finish with the Lightning Round. Do you drive a
customized car?
Speaker 1 (46:24):
I don't actually yet. Well, if I am, what have I.
Speaker 2 (46:27):
Seen on your Instagram? What's that truck you keep posting
on your Instagram?
Speaker 1 (46:31):
So I so I have a truck that's customized. I
don't drive it around as much, but maybe this year
I'll start taking it out. This year I've been you know,
we have been kind of stelf about it, talking about it,
but we haven't talked about it in a big way
because we have a big release coming soon.
Speaker 2 (46:45):
I mean, you're literally posting it on Instagram. It's not
that stuff. Tell me about Tell me about that truck
you keep posting on Instagram? What's going on with that?
Speaker 1 (46:53):
So it's a truck is fully the full body is
fully customized.
Speaker 2 (46:56):
It says anvil in the back when you post it?
Is it called anvil?
Speaker 1 (46:59):
Dumb question and call it and call it anvil? I
think it's the idea was actually the shape design of
it was inspired by amil. If you look at the
front fender, it actually looks like a the front bumper
looks like an anvil. But also the idea is that, like,
you know, we're actually forming shits on an anvil. So yeah,
it was very fitting.
Speaker 2 (47:15):
Tell me about that truck, like, tell just tell me
what's it look like.
Speaker 1 (47:18):
Yeah, so for example, like you know, we put a
lot of form and sharp edges in the in the hood. Right.
Most vehicles have a very hard time if you look
at most of the you know, hood of the vehicles,
they are very smooth because it's very hard to actually
put sharp angles in the hood. So if you look
at this truck, this truck has a lot of angles,
a lot of sharp detail right in the hood. Right,
(47:41):
And and and that's very expressive of the type of
person for example, that I am, right, I like things
that are edgy, and and that truck is certainly edgy. Right.
It's bare metal, right, you know, there is no blemishes
being hiden and hidden under the under the vehicle. You know,
a lot of people when cyber Truck came out, we
(48:02):
got very excited about you know, oh, it's bare metal.
It looks like a metal, but then there was no
form in it because it's actually very hard to make
it for metal look nice. And so that's one of
the things we wanted to show. We want to show that, Okay,
you can actually have a form metal with a lot
of detail in it and still keep it bare metal
because it will look nice. Right, So, yeah, a lot
(48:22):
of design features of it. For me kind of represents
the type of personality and character that I have. But
I think that's how every car should be. You know,
people should be able to have that freedom to choose
what their cars look like.
Speaker 2 (48:34):
How many skull tattoos do you have?
Speaker 1 (48:38):
I've got three? Why it's it's so, yeah, it's an
interesting thing. So a skull for me represents kind of
and it's an abstract for death of ego. So I
have a tattoo on my thumb which is a skull
that's holding a microphone to his ears. And this was
(49:01):
a time where you know, I felt like, you know,
I had a good platform and I could talk a
lot and people would listen. But then I realized I
should yes, that's right, but I should maybe keep them
keep the mic close to my ears and also listen
as opposed to talk all the time, right, So I think.
Speaker 2 (49:17):
Skull microphones don't work that way for the record, but.
Speaker 1 (49:21):
I like it as a metaphor exactly. But I think
the idea of is around, you know, kind of reminders
of you can see a lot of my tattoos on
my on my hands, so it's a really reminder for
myself to know that, you know, be present and make
sure that you know you're not involved with your ego
too much and you can see others people's perspective.
Speaker 2 (49:41):
Is there any tension between ego death and custom cars?
Speaker 1 (49:46):
Tension between ego death and custom cars? I don't know.
Speaker 2 (49:50):
I'm just playing, but like you know, custom car kind
of seems like, hey, look at me, I'm special, and
ego death seems like, oh, don't look at me, I'm
not so special.
Speaker 1 (49:59):
Yeah, no, I think I think the difference is I think, yeah,
if you have attachment to your custom card, then maybe
there's tension. But I more think of it in terms
of expression. Right. You know, you can be an artist,
you can. You can you can design your home the
way it expresses you. You can design the theme of
your podcast the way it expresses you. You can design
your car. Also, the way it expresses you. I think
(50:21):
it's leus so about oh look at me, I'm special.
It's more like, here's my expression to the world for
the people to see. But I think that expressiveness is
it is pretty amazing. I think that's uniquely one of
the unique things about humans that like, you know, we
we I think all we do when we come to
this world is expressing ourselves right, expressing uself through our work,
expressing through ourselves through our relationships. And if you can
(50:45):
enable people to express themselves better better, I think that's great.
But if you get attached to your expressions and your
ideas and your thoughts and think, oh, I'm better than
everybody else, and I think that that becomes that becomes
a little bit of an ego driven trip.
Speaker 2 (51:05):
Edward Mayer is the co founder and CEO of Mocking Labs.
Today's show was produced by Gabriel Hunter Cheang. It was
edited by Lyddy jeene Kott and engineered by Sarah Bruginner.
You can email us at problem at Pushkin dot FM.
I'm Jacob Goldstein and we'll be back next week with
another episode of What's Your Problem.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.
True Crime Tonight
If you eat, sleep, and breathe true crime, TRUE CRIME TONIGHT is serving up your nightly fix. Five nights a week, KT STUDIOS & iHEART RADIO invite listeners to pull up a seat for an unfiltered look at the biggest cases making headlines, celebrity scandals, and the trials everyone is watching. With a mix of expert analysis, hot takes, and listener call-ins, TRUE CRIME TONIGHT goes beyond the headlines to uncover the twists, turns, and unanswered questions that keep us all obsessed—because, at TRUE CRIME TONIGHT, there’s a seat for everyone. Whether breaking down crime scene forensics, scrutinizing serial killers, or debating the most binge-worthy true crime docs, True Crime Tonight is the fresh, fast-paced, and slightly addictive home for true crime lovers.