1
00:00:00,000 --> 00:00:03,661
Edward Mehr: these are like a 70
year old plane. So they need to

2
00:00:03,661 --> 00:00:06,921
have the dies for it, the
tooling for it, and the tooling

3
00:00:06,921 --> 00:00:10,582
also sometimes doesn't exist. So
for example, for a landing gear

4
00:00:10,582 --> 00:00:13,957
door of a certain aircraft,
we're looking at four year lead

5
00:00:13,957 --> 00:00:17,446
time and millions of dollars
before you can get your part. so

6
00:00:17,446 --> 00:00:21,107
the aircraft needs to be down on
the ground for four years. it's

7
00:00:21,107 --> 00:00:24,710
a huge, hugely affect the fleet
readiness, for our military. so

8
00:00:24,710 --> 00:00:28,314
we're working with them turning
some of those four yearly times

9
00:00:28,314 --> 00:00:29,000
into days. ​

10
00:00:29,000 --> 00:00:33,288
Audrow Nash: There's a big
interest in America to reshore

11
00:00:33,288 --> 00:00:37,952
manufacturing. Now, to do this,
do we just bring the jobs back

12
00:00:37,952 --> 00:00:42,617
to the U. S. and overlook that
the reasons that those jobs may

13
00:00:42,617 --> 00:00:47,206
have moved may still exist? Or
do we invent new ways of doing

14
00:00:47,206 --> 00:00:51,042
things and leverage our
strengths in technology and

15
00:00:51,042 --> 00:00:55,330
innovation. This interview is
with Edward Mehr, and he is

16
00:00:55,330 --> 00:00:59,844
firmly in the second camp. He
wants to invent new ways to do

17
00:00:59,844 --> 00:01:03,982
manufacturing, and he's well
positioned to help in this

18
00:01:03,982 --> 00:01:08,345
effort. He's the CEO and a co
founder of Machina Labs, and

19
00:01:08,345 --> 00:01:12,708
they're reinventing metal
bending with robotics and AI. In

20
00:01:12,708 --> 00:01:16,921
some cases, taking something
that literally is taking us

21
00:01:16,921 --> 00:01:21,360
years to make, and they're doing
it in just days. It was an

22
00:01:21,360 --> 00:01:26,024
awesome conversation and got me
excited about the potential of

23
00:01:26,024 --> 00:01:30,839
robots and AI to greatly improve
how we do manufacturing. You'll

24
00:01:30,839 --> 00:01:35,503
like this interview if you're
curious about robotics and AI in

25
00:01:35,503 --> 00:01:40,092
manufacturing, interested in how
great robotics companies are

26
00:01:40,092 --> 00:01:44,455
built, and if you're interested
in AI for modeling complex

27
00:01:44,455 --> 00:01:48,969
phenomena like metal bending.
Without further ado, here's my

28
00:01:48,969 --> 00:01:51,000
conversation with Edward. ​

29
00:01:51,000 --> 00:01:53,000
Audrow Nash: Would you introduce
yourself?

30
00:01:53,000 --> 00:01:56,519
Edward Mehr: Yes, my name is
Edward Mehr. I am CEO and co

31
00:01:56,519 --> 00:01:58,000
founder at Machina Labs.

32
00:01:58,000 --> 00:02:01,000
Audrow Nash: Hell yeah. Would
you tell me about Machina Labs?

33
00:02:01,000 --> 00:02:05,062
Edward Mehr: Yeah, so we are,
working on the next generation

34
00:02:05,062 --> 00:02:08,921
of manufacturing floors. the
real problem we're trying to

35
00:02:08,921 --> 00:02:13,118
solve is that today, if you want
to build a physical part, you

36
00:02:13,118 --> 00:02:16,368
pretty much have to build a
factory that is very

37
00:02:16,368 --> 00:02:20,429
specifically built for that
part. A lot of tooling, a lot of

38
00:02:20,429 --> 00:02:24,288
machinery that goes into
factories, are very specifically

39
00:02:24,288 --> 00:02:28,350
designed for the geometry, for
the material you're trying to

40
00:02:28,350 --> 00:02:32,615
manufacture. And that severely
limits in terms of the amount of

41
00:02:32,615 --> 00:02:36,000
CapEx investment you have to
make to build a part,

42
00:02:36,000 --> 00:02:39,052
Audrow Nash: So it's far more
expensive because you have to

43
00:02:39,052 --> 00:02:42,000
invest in infrastructure to
build it. Okay. Or equipment.

44
00:02:42,000 --> 00:02:45,392
Edward Mehr: What we're trying
to do is, can we develop

45
00:02:45,392 --> 00:02:48,969
technologies that allow you to,
make a part A today out of

46
00:02:48,969 --> 00:02:52,423
design A and material B, and
then switch to design C and

47
00:02:52,423 --> 00:02:56,000
material D tomorrow without
having to change your factory.

48
00:02:56,000 --> 00:02:59,836
Audrow Nash: I love it. Hell
yeah. So how, are you

49
00:02:59,836 --> 00:03:04,517
approaching that? Because that,
sounds very interesting to be

50
00:03:04,517 --> 00:03:09,350
able to switch so quickly. And
I've heard that we're going from

51
00:03:09,350 --> 00:03:14,107
like high volume low mix meaning
a lot of the same thing being

52
00:03:14,107 --> 00:03:18,558
made to more of lower volume
with high mix and this sounds

53
00:03:18,558 --> 00:03:23,391
like it's working towards that
what are some like what do these

54
00:03:23,391 --> 00:03:26,000
manufacturing techniques look
like

55
00:03:26,000 --> 00:03:29,801
Edward Mehr: Yeah, it's funny
you mentioned it going from high

56
00:03:29,801 --> 00:03:33,418
volume low mix to high mix low
volume. I think actually the

57
00:03:33,418 --> 00:03:36,973
ideal combination is high mix,
high volume, right? because

58
00:03:36,973 --> 00:03:40,590
always people think about it in
terms of it's a compromise,

59
00:03:40,590 --> 00:03:42,000
right? Because that was

60
00:03:42,000 --> 00:03:43,000
Audrow Nash: a trade off

61
00:03:43,000 --> 00:03:46,464
Edward Mehr: was given as a
trade off, right? Like you need

62
00:03:46,464 --> 00:03:50,221
to have high volume and low mix
or low mix and high volume. What

63
00:03:50,221 --> 00:03:53,860
is it? But If we want to think
about it fundamentally, what is

64
00:03:53,860 --> 00:03:57,441
the technology that needs to be
there to enable both high and

65
00:03:57,441 --> 00:04:00,905
low volume and high and low mix?
that's actually what we're

66
00:04:00,905 --> 00:04:04,427
thinking about. but to answer
your question directly, if you

67
00:04:04,427 --> 00:04:07,186
look at history of
manufacturing, and if you go

68
00:04:07,186 --> 00:04:10,943
Maybe 200, 300 years ago, where
manufacturing first started as a

69
00:04:10,943 --> 00:04:14,641
craft, not to start it, but like
it was mostly done as a craft.

70
00:04:14,641 --> 00:04:17,576
We actually had a lot of
flexibility. We had these

71
00:04:17,576 --> 00:04:21,157
people, we call them craftsmen
or blacksmiths. and they had a

72
00:04:21,157 --> 00:04:24,621
very creative mind. They had
very limited set of tools like

73
00:04:24,621 --> 00:04:28,319
hammers and chisels, maybe a set
of tool of tools. 10 different

74
00:04:28,319 --> 00:04:32,076
kind, but then you could one day
go to them and say, Hey, I have

75
00:04:32,076 --> 00:04:35,716
this piece of rod, can you turn
it into a sword? And then they

76
00:04:35,716 --> 00:04:39,355
would hammer it into a sword,
right? They will figure out with

77
00:04:39,355 --> 00:04:43,112
their mind, how they're going to
use their limited tool set they

78
00:04:43,112 --> 00:04:46,752
have to make a, sword. And the
next day you can go to them and

79
00:04:46,752 --> 00:04:50,215
be like, I have this piece of
sheet of metal, and I want to

80
00:04:50,215 --> 00:04:53,679
make a shield. And then they
would use the exact same tools

81
00:04:53,679 --> 00:04:57,377
to use the sword, but apply it
differently in a creative way to

82
00:04:57,377 --> 00:05:00,841
make you a shield. So they were
actually very flexible. The

83
00:05:00,841 --> 00:05:04,304
challenge was that a craftsman
was a very much of a learned

84
00:05:04,304 --> 00:05:08,061
skill. You had to do mentorship.
You had to, to go learn it from

85
00:05:08,061 --> 00:05:11,701
somebody else who has done it.
so it would take a while before

86
00:05:11,701 --> 00:05:15,340
somebody would become expert in
it. And then more importantly,

87
00:05:15,340 --> 00:05:19,156
humans, have limitation in terms
of throughput. So you could make

88
00:05:19,156 --> 00:05:22,678
you maybe make one sword, two
sword, three swords a day. but

89
00:05:22,678 --> 00:05:25,731
beyond that, You couldn't,
right? So with industrial

90
00:05:25,731 --> 00:05:29,077
revolution, there's a change
happened where we were like,

91
00:05:29,077 --> 00:05:32,482
okay, craftsmen are great,
flexible, but they can't have a

92
00:05:32,482 --> 00:05:36,063
throughput. So we build these
machines that can make the same

93
00:05:36,063 --> 00:05:39,174
thing over and over again,
right? Because we were not

94
00:05:39,174 --> 00:05:42,872
creative enough or didn't have
enough technology to develop the

95
00:05:42,872 --> 00:05:46,042
same thing that the craftsman
does. We could make very

96
00:05:46,042 --> 00:05:49,623
constraint machines, hardware
constraint machines that can do

97
00:05:49,623 --> 00:05:52,911
the same thing over and over
again. but we couldn't make

98
00:05:52,911 --> 00:05:56,609
flexibility because we couldn't
Replicate intelligence that the

99
00:05:56,609 --> 00:06:00,072
craftsman has. So for the past
two centuries or more, up to

100
00:06:00,072 --> 00:06:03,477
even today, manufacturing
morphed into make the same thing

101
00:06:03,477 --> 00:06:07,234
over and over again. Even today,
if you're like, for example, an

102
00:06:07,234 --> 00:06:10,580
automotive business, a car
business, the main way you can

103
00:06:10,580 --> 00:06:14,044
make money is create one car
that everybody loves. And then

104
00:06:14,044 --> 00:06:17,390
make a lot of it. That's how you
can make margin. because

105
00:06:17,390 --> 00:06:20,736
manufacturing is very much
dependent on the type of parts

106
00:06:20,736 --> 00:06:24,317
you're making and you cannot
easily change it. Look at Tesla,

107
00:06:24,317 --> 00:06:27,546
they're planning to make 5, 6
million of Model Y a year

108
00:06:27,546 --> 00:06:31,068
sometimes. sometimes. and that
means that, the game is still

109
00:06:31,068 --> 00:06:34,473
the same as when industrial
revolution started, right? You

110
00:06:34,473 --> 00:06:37,995
have to make the same thing over
and over again, to make it,

111
00:06:37,995 --> 00:06:41,224
profitable. Now, today.
Advancement in AI and Robotics.

112
00:06:41,224 --> 00:06:44,511
We could replicate what a
craftsman does at a scale. You

113
00:06:44,511 --> 00:06:47,681
have robotics that has the same
kinematic freedom as a

114
00:06:47,681 --> 00:06:50,910
craftsman, right? Maybe even
more dexterity, maybe more

115
00:06:50,910 --> 00:06:54,314
precision, higher force that
they can apply. But then with

116
00:06:54,314 --> 00:06:57,661
AI, we also can replicate what
happens in the mind of the

117
00:06:57,661 --> 00:07:00,889
craftsman. We can figure out how
to use different tools

118
00:07:00,889 --> 00:07:04,177
creatively to replicate
different types of processes. So

119
00:07:04,177 --> 00:07:07,523
I think now we're, the two
ingredients To make a flexible

120
00:07:07,523 --> 00:07:10,693
manufacturing system exist. So
to answer your question

121
00:07:10,693 --> 00:07:13,569
directly, our systems use
robotics and artificial

122
00:07:13,569 --> 00:07:17,326
intelligence to almost replicate
what a craftsman does, but in a

123
00:07:17,326 --> 00:07:20,966
scalable way. Because once you
build one robotic system can do

124
00:07:20,966 --> 00:07:24,194
a craftsman, I can just
replicate it and have thousands

125
00:07:24,194 --> 00:07:27,717
and thousands of those systems
that does the same thing. Now

126
00:07:27,717 --> 00:07:31,239
you get flexibility, but you
also get scale through, through

127
00:07:31,239 --> 00:07:33,000
scaling these robotic systems.

128
00:07:33,000 --> 00:07:37,782
Audrow Nash: Okay. So I'm
interested in what you are

129
00:07:37,782 --> 00:07:43,115
referring to with AI for this.
what are you, what kinds of

130
00:07:43,115 --> 00:07:48,908
things can we use AI to come in
and help with manufacturing and

131
00:07:48,908 --> 00:07:54,701
maybe some simple reasoning or
about what to do next or what to

132
00:07:54,701 --> 00:07:57,000
do? how are you using ai?

133
00:07:57,000 --> 00:08:00,000
Edward Mehr: AI, the definition
of AI has changed over time

134
00:08:00,000 --> 00:08:01,000
Audrow Nash: Oh, definitely.

135
00:08:01,000 --> 00:08:04,612
Edward Mehr: Since its
inception. If you look at 1960s,

136
00:08:04,612 --> 00:08:08,421
and I come from a computer
science background, if you look

137
00:08:08,421 --> 00:08:12,164
at 1960s, intelligent AI systems
were actually rule based

138
00:08:12,164 --> 00:08:16,236
systems. We'll call rule based
systems intelligent AI systems,

139
00:08:16,236 --> 00:08:20,110
like regular programming. And
then over time, as we develop

140
00:08:20,110 --> 00:08:23,985
new techniques, new empirical
techniques, where we can find

141
00:08:23,985 --> 00:08:28,056
patterns in the data and, use
those patterns to actually do an

142
00:08:28,056 --> 00:08:31,734
object, do a task that a human
does more easily and more

143
00:08:31,734 --> 00:08:35,477
automated. Then it starts
shaping the definition of AI to

144
00:08:35,477 --> 00:08:39,549
more things that the humans can
do. there are a whole array of

145
00:08:39,549 --> 00:08:42,832
tasks in manufacturing that
falls within different

146
00:08:42,832 --> 00:08:46,969
categories. Up to today, most of
the automation has been what I

147
00:08:46,969 --> 00:08:50,778
call classic, classical AI,
where you define the rules and

148
00:08:50,778 --> 00:08:54,587
the system just follows the
rules. if you look at the type

149
00:08:54,587 --> 00:08:58,199
of manufacturing we talked
about, where a craftsman can

150
00:08:58,199 --> 00:09:02,468
figure out a sequence of events.
can look at a piece of metal and

151
00:09:02,468 --> 00:09:06,605
as it's hammering it, figure out
what is the next steps that it

152
00:09:06,605 --> 00:09:10,808
needs to do. That's the type of
AI we're talking about. It's not

153
00:09:10,808 --> 00:09:14,682
following a set of rules, it's
finding patterns in making a

154
00:09:14,682 --> 00:09:18,820
task done. And that task might
be forming a sheet of metal into

155
00:09:18,820 --> 00:09:22,563
precision from a flat sheet.
What are the set of, process

156
00:09:22,563 --> 00:09:26,766
parameters? What are the set of,
sequence of tasks that needs to

157
00:09:26,766 --> 00:09:30,772
be done to get a flat sheet of
metal all the way to the final

158
00:09:30,772 --> 00:09:34,646
product? You cannot easily
define it. You cannot define it.

159
00:09:34,646 --> 00:09:38,915
It's not a rule base. That's why
it has been a craft. You have to

160
00:09:38,915 --> 00:09:42,855
learn it from somebody, mentor.
but with AI, we can actually

161
00:09:42,855 --> 00:09:47,058
capture what is happening in the
process using different ways of

162
00:09:47,058 --> 00:09:50,932
capturing data. And then it
starts identifying what are the

163
00:09:50,932 --> 00:09:55,004
patterns that the craftsman uses
to get to the right part. And

164
00:09:55,004 --> 00:09:59,010
that's what we're talking about
here. it's similar to some of

165
00:09:59,010 --> 00:10:03,016
the things you might see in, in
the, in. Discussions that you

166
00:10:03,016 --> 00:10:06,759
see around AI outside of
manufacturing, how can a ChatGPT

167
00:10:06,759 --> 00:10:10,502
then do similar to human
reasoning. That's the type of AI

168
00:10:10,502 --> 00:10:14,639
we're talking about here. That
being said, we also use a lot of

169
00:10:14,639 --> 00:10:18,251
more earlier versions of AI,
where, for example, as I'm

170
00:10:18,251 --> 00:10:22,126
forming a part, can I look at
the Data from the sensors and

171
00:10:22,126 --> 00:10:26,394
find an anomalous behavior where
it might lead into a defect into

172
00:10:26,394 --> 00:10:30,532
the part. So classical, anomaly
detection using empirical data.

173
00:10:30,532 --> 00:10:34,603
Those are the things that I also
use. But the main enabler is,

174
00:10:34,603 --> 00:10:38,478
figuring out how a craftsman
forms a part or crafts a part.

175
00:10:38,478 --> 00:10:42,221
What are the set of rules and,
guidelines that we need to

176
00:10:42,221 --> 00:10:45,899
develop so that the robotic
system can actually have the

177
00:10:45,899 --> 00:10:48,000
same performance as a craftsman?

178
00:10:48,000 --> 00:10:51,736
Audrow Nash: Okay. I would love
if you can take me through a

179
00:10:51,736 --> 00:10:55,286
concrete example, because I,
feel like I understand a bit

180
00:10:55,286 --> 00:10:58,897
from how you're describing it,
but I want to see like in a

181
00:10:58,897 --> 00:11:02,696
metal forming or something like
this, how do you, how are you

182
00:11:02,696 --> 00:11:05,000
using this to achieve a desired
part?

183
00:11:05,000 --> 00:11:08,677
Edward Mehr: Yes. So for folks
who might have seen our videos,

184
00:11:08,677 --> 00:11:11,998
you will see two robots, for
example, forming a sheet of

185
00:11:11,998 --> 00:11:15,438
metal from flat sheet. So you
have a flat sheet, two giant

186
00:11:15,438 --> 00:11:19,174
robots on two sides, and almost
like a potter, they're like two

187
00:11:19,174 --> 00:11:22,910
stylists looking endeffectors on
two robots on two sided sheet.

188
00:11:22,910 --> 00:11:26,291
They come in and pinch and
deform a sheet of metal slowly

189
00:11:26,291 --> 00:11:29,849
into shape. Now, when you look
at it, it almost looks like a

190
00:11:29,849 --> 00:11:33,467
very much like a CNC process,
right? Where you're almost like

191
00:11:33,467 --> 00:11:37,322
doing waterlines of the part and
deforming the part in waterlines

192
00:11:37,322 --> 00:11:40,939
until you get the final part.
But in reality, it's not just a

193
00:11:40,939 --> 00:11:44,438
CNC process. It's a lot of
different changes in the process

194
00:11:44,438 --> 00:11:48,115
parameters, but also the path
the robot takes is sometimes you

195
00:11:48,115 --> 00:11:51,437
might go in the center of the
part and start forming the

196
00:11:51,437 --> 00:11:55,232
center of the part, and then go
back to the outside and form the

197
00:11:55,232 --> 00:11:58,791
outside of the part. So why do
we have to do that? Why can't

198
00:11:58,791 --> 00:12:02,586
just a simple heuristic CNC path
doesn't get you the right part?

199
00:12:02,586 --> 00:12:06,144
It's because, as you're forming
the part, the sheet moves in

200
00:12:06,144 --> 00:12:09,584
unpredictable ways. And also,
there's a process, there's a

201
00:12:09,584 --> 00:12:13,321
phenomenon called a springback,
where you might form a sheet to

202
00:12:13,321 --> 00:12:16,642
a certain length or certain
depth. After the robot moves

203
00:12:16,642 --> 00:12:20,259
away, the sheet jumps back into
a different shape, right? you

204
00:12:20,259 --> 00:12:23,877
need to account for all these
movements, all these springback

205
00:12:23,877 --> 00:12:27,613
phenomenons that we talked about
to form the right part. So you

206
00:12:27,613 --> 00:12:31,231
end up, the path that the robot
ends up taking to form a part

207
00:12:31,231 --> 00:12:34,196
that's like less than a
millimeter accurate to the

208
00:12:34,196 --> 00:12:37,755
design is a very non intuitive
path. You can, a human cannot

209
00:12:37,755 --> 00:12:39,000
just think of it. and

210
00:12:39,000 --> 00:12:41,652
Audrow Nash: you can't think of
it all at once, but you can

211
00:12:41,652 --> 00:12:43,000
probably iterate to get there.

212
00:12:43,000 --> 00:12:46,363
Edward Mehr: you could iterate
to get there. Yes. And that's

213
00:12:46,363 --> 00:12:49,671
what, we actually generate the
data. So for right now, what

214
00:12:49,671 --> 00:12:52,922
happens is, you run maybe a
heuristic path that you think,

215
00:12:52,922 --> 00:12:56,342
oh, this might work well. And
then you scan it and say, okay,

216
00:12:56,342 --> 00:12:59,705
This area was five inches off,
that area was two inches off,

217
00:12:59,705 --> 00:13:03,181
compared to the CAD. So I can
make a slight adjustments to get

218
00:13:03,181 --> 00:13:06,712
it a little slightly closer, and
then run another one, and then

219
00:13:06,712 --> 00:13:10,356
see the results, and run another
one. But if I could have a model

220
00:13:10,356 --> 00:13:13,776
that would tell me, in order to
get to this part, this is the

221
00:13:13,776 --> 00:13:17,363
non intuitive path you can take.
And you didn't have to do those

222
00:13:17,363 --> 00:13:20,727
trials. you can significantly
improve the performance of the

223
00:13:20,727 --> 00:13:24,146
system. So what we're doing at
Machina is like we're starting

224
00:13:24,146 --> 00:13:27,061
with human intelligence and
creating a very seamless

225
00:13:27,061 --> 00:13:30,145
interface that humans can
iterate and build these parts

226
00:13:30,145 --> 00:13:33,508
while we capture this data and
use this data to build models

227
00:13:33,508 --> 00:13:36,423
that slowly improve the
efficiency of the humans. in

228
00:13:36,423 --> 00:13:39,618
early days, we would do 25
trials to get a part done. Now

229
00:13:39,618 --> 00:13:43,262
we are down to five, six trials.
And the goal is get to the point

230
00:13:43,262 --> 00:13:45,000
that right off the bat, you get

231
00:13:45,000 --> 00:13:46,000
Audrow Nash: Goes to the right
thing.

232
00:13:46,000 --> 00:13:48,000
Edward Mehr: done. Yeah.

233
00:13:48,000 --> 00:13:52,492
Audrow Nash: That's really cool.
Hell yeah. So the way I'm

234
00:13:52,492 --> 00:13:57,139
understanding it is you, so in
this metal forming task where

235
00:13:57,139 --> 00:14:01,786
you have the two robots on the
one on each side of the sheet

236
00:14:01,786 --> 00:14:06,587
and they pinch so that they can
push in to deform the metal in

237
00:14:06,587 --> 00:14:11,157
some sort of way to make some
complex part, you have a part

238
00:14:11,157 --> 00:14:15,803
that you want to make and you
are iterating in some sense so

239
00:14:15,803 --> 00:14:20,373
you like you'll push you'll
pinch in some way and deform it

240
00:14:20,373 --> 00:14:25,252
and you do that a few times and
you say this is what I expected

241
00:14:25,252 --> 00:14:29,511
this is what it actually looks
like when you scan it or

242
00:14:29,511 --> 00:14:34,003
something to understand the
state of the part and then you

243
00:14:34,003 --> 00:14:38,186
can keep steps until you
converge on what you actually

244
00:14:38,186 --> 00:14:42,987
want the part to look like then
you're also generating data to

245
00:14:42,987 --> 00:14:47,634
make it so you have a better,
you're generating data so that

246
00:14:47,634 --> 00:14:52,049
you can use that to have a
better model of how all of the

247
00:14:52,049 --> 00:14:57,005
things you're doing to the metal
sheet or whatever the sheet is,

248
00:14:57,005 --> 00:15:01,497
you're understanding what
happens when you apply an action

249
00:15:01,497 --> 00:15:06,066
to that sheet so that you could
be more efficient with your

250
00:15:06,066 --> 00:15:10,946
actions, Then you're more likely
to be reliable and then you're

251
00:15:10,946 --> 00:15:15,515
also faster because you are
better able to make the part in

252
00:15:15,515 --> 00:15:19,000
fewer actions and this kind of
thing is that?

253
00:15:19,000 --> 00:15:20,000
Edward Mehr: No, you're exactly,

254
00:15:19,000 --> 00:15:22,439
Edward Mehr: you're basically
building models that gives you

255
00:15:22,439 --> 00:15:25,707
an understanding of what is the
physics that is happening

256
00:15:25,707 --> 00:15:28,000
underneath. And use the data to
do that.

257
00:15:28,000 --> 00:15:31,144
Audrow Nash: Rather than
directly modeling it because

258
00:15:31,144 --> 00:15:34,644
it's very complex because of all
the really complex contact

259
00:15:34,644 --> 00:15:35,000
forces

260
00:15:35,000 --> 00:15:38,389
Edward Mehr: Contact forces,
friction, so if you want to,

261
00:15:38,389 --> 00:15:41,957
traditionally people when they
would do this, They would use

262
00:15:41,957 --> 00:15:45,584
physics based modeling, right?
people think of finite element

263
00:15:45,584 --> 00:15:49,389
analysis or computational fluid
dynamics. Those are methods that

264
00:15:49,389 --> 00:15:52,303
allows you to simulate a
physical phenomenon. The

265
00:15:52,303 --> 00:15:55,989
challenge is that with a lot of
these processes, first of all,

266
00:15:55,989 --> 00:15:57,000
FEA is very slow,

267
00:15:57,000 --> 00:15:58,000
Audrow Nash: What was that?

268
00:15:58,000 --> 00:16:00,351
Edward Mehr: FEA or Finite
Element Analysis, which is the

269
00:16:00,351 --> 00:16:02,000
physics based way of doing these
things.

270
00:16:02,000 --> 00:16:04,929
Audrow Nash: Yeah, you simulate
like a bunch of points and you

271
00:16:04,929 --> 00:16:07,764
see what happens and you do,
it's a differential equation to

272
00:16:07,764 --> 00:16:08,000
solve

273
00:16:08,000 --> 00:16:11,405
Edward Mehr: You apply physical
laws to, to figure out, okay,

274
00:16:11,405 --> 00:16:15,033
what is happening. One challenge
with that approach is that it is

275
00:16:15,033 --> 00:16:18,605
a slow. So it requires a lot of
computation to, to do it. We, in

276
00:16:18,605 --> 00:16:21,786
early days of our company, we
would form these parts that

277
00:16:21,786 --> 00:16:25,079
would like to take 20 minutes to
form. And if you wanted to

278
00:16:25,079 --> 00:16:28,540
simulate it using physics based
models, it would take us, on a

279
00:16:28,540 --> 00:16:32,000
27 core machine, it would take
us a week. So we're like, okay,

280
00:16:32,000 --> 00:16:33,000
Audrow Nash: Not feasible. Yeah.

281
00:16:33,000 --> 00:16:36,000
Edward Mehr: Not feasible. I
would rather just run the part,

282
00:16:36,000 --> 00:16:37,000
Audrow Nash: Yeah.

283
00:16:36,000 --> 00:16:38,458
Edward Mehr: Let's see the
results in real life, as opposed

284
00:16:38,458 --> 00:16:40,000
to simulate. Simulate it with
nature.

285
00:16:40,000 --> 00:16:43,131
Audrow Nash: like heat people's
homes with, computers that are

286
00:16:43,131 --> 00:16:45,000
running at full, bore the whole
time.

287
00:16:45,000 --> 00:16:47,926
Edward Mehr: So we, made it, we
changed the parts. We're like,

288
00:16:47,926 --> 00:16:50,710
okay, maybe can we do this,
build these models empirically?

289
00:16:50,710 --> 00:16:53,399
And the other challenge with
physics based models is that

290
00:16:53,399 --> 00:16:56,419
even if they're fast, There are
not accurate. To your point, you

291
00:16:56,419 --> 00:16:59,345
might not be modeling everything
that's in there. For example,

292
00:16:59,345 --> 00:17:01,704
there might be some
miscalibration in a robot that

293
00:17:01,704 --> 00:17:04,630
you're not modeling. There might
be some, friction forces that

294
00:17:04,630 --> 00:17:07,650
causes some issues. There might
be some adhesion between the end

295
00:17:07,650 --> 00:17:10,387
effector and the material that
you're not modeling, or you

296
00:17:10,387 --> 00:17:11,000
don't, you're

297
00:17:11,000 --> 00:17:14,697
Audrow Nash: The model just
isn't good enough. To do the

298
00:17:14,697 --> 00:17:18,658
task. like one thing that really
struck me, in my electrical

299
00:17:18,658 --> 00:17:22,553
engineering education is so
voltage equals resistance times

300
00:17:22,553 --> 00:17:26,581
current, but so V equals IR, but
not at high frequencies. and

301
00:17:26,581 --> 00:17:30,542
then so the model breaks down.
So the model is correct up to

302
00:17:30,542 --> 00:17:34,833
some, in some environment, or at
least pretty close to correct in

303
00:17:34,833 --> 00:17:38,794
some environment. And so if
you're doing complex things on a

304
00:17:38,794 --> 00:17:42,557
part, I would imagine the
physics based models just don't

305
00:17:42,557 --> 00:17:45,000
capture all the phenomenon very
well.

306
00:17:45,000 --> 00:17:47,684
Edward Mehr: Yeah, you're
exactly right. And that's once

307
00:17:47,684 --> 00:17:50,655
there's a need to build these
empirical models, they're faster

308
00:17:50,655 --> 00:17:53,723
to inference, right? So you can
get the result of it really fast

309
00:17:53,723 --> 00:17:56,215
once you build it. And then
because they're based on

310
00:17:56,215 --> 00:17:59,043
physical data, you're almost
taking into account everything

311
00:17:59,043 --> 00:18:01,918
that's happening in real world.
The key though is, as you're

312
00:18:01,918 --> 00:18:04,938
building these robotic systems,
you need to build a system that

313
00:18:04,938 --> 00:18:07,909
allows you to expose that data
you need to build these models.

314
00:18:07,909 --> 00:18:10,689
And that's what some of the
architectural decisions we had

315
00:18:10,689 --> 00:18:13,469
to make is that, how can I
capture the data that I need at

316
00:18:13,469 --> 00:18:16,201
every millisecond of this
process? So I can model it. and

317
00:18:16,201 --> 00:18:19,172
affects your product development
and what kind of robotics you

318
00:18:19,172 --> 00:18:22,000
want to use and what kind of
control loops you want to use.

319
00:18:22,000 --> 00:18:26,820
Audrow Nash: Okay. So I just,
that is very interesting. I want

320
00:18:26,820 --> 00:18:31,796
to get more into that, but just
so I understand at a high level,

321
00:18:31,796 --> 00:18:36,772
as I'm understanding things now,
you have two parts, really. one

322
00:18:36,772 --> 00:18:41,359
is the iterative working towards
a task using your existing

323
00:18:41,359 --> 00:18:46,413
model. And the other one is the,
I've captured data. Let me learn

324
00:18:46,413 --> 00:18:51,000
a better model part. So you have
those two things. Okay. so

325
00:18:51,000 --> 00:18:54,348
Audrow Nash: how did you design
it for, how did you design your

326
00:18:54,348 --> 00:18:57,643
system so that you can make an
efficient transfer to your data

327
00:18:57,643 --> 00:19:00,725
part of this so that you can
learn a better model? Because

328
00:19:00,725 --> 00:19:02,000
that's very interesting.

329
00:19:02,000 --> 00:19:05,765
Edward Mehr: Yeah, so there is,
the choice of your architecture,

330
00:19:05,765 --> 00:19:09,235
your hardware architecture
greatly affects your ability to,

331
00:19:09,235 --> 00:19:12,529
get data. So you can imagine,
for example, in a stamping

332
00:19:12,529 --> 00:19:15,706
operation, which is a
traditional way of forming sheet

333
00:19:15,706 --> 00:19:19,235
metal parts, where you have a
giant press, and you have dyes

334
00:19:19,235 --> 00:19:22,882
that are male and a female dye,
and you put a sheet in between

335
00:19:22,882 --> 00:19:26,471
and you stamp it. Not a whole
lot of opportunities to capture

336
00:19:26,471 --> 00:19:28,000
data, okay, we're going to

337
00:19:28,000 --> 00:19:29,000
Audrow Nash: it's like start and
end

338
00:19:29,000 --> 00:19:32,606
Edward Mehr: It's a start and
then you can capture that Some

339
00:19:32,606 --> 00:19:36,391
kind of an input to the to the
either it's a hydraulic press or

340
00:19:36,391 --> 00:19:39,937
a servo press the current or
hydraulic pressure Or whatever

341
00:19:39,937 --> 00:19:43,482
throughout the time, but you
don't really you cannot really

342
00:19:43,482 --> 00:19:46,968
know what is happening to your
part in a very granular way

343
00:19:46,968 --> 00:19:50,333
Because it's inside a very
destructive stamping press so

344
00:19:50,333 --> 00:19:54,059
with our process because we're
incrementally forming it we can

345
00:19:54,059 --> 00:19:57,844
capture A lot of little details,
like what is the forces, every

346
00:19:57,844 --> 00:20:01,630
once in a while, scan it, what
is the response, on both robots,

347
00:20:01,630 --> 00:20:05,176
how much current we're putting
through the motors, how much

348
00:20:05,176 --> 00:20:08,000
deflection we're looking at each
of the robots,

349
00:20:08,000 --> 00:20:11,482
Audrow Nash: wow, so you're
doing this on the full system.

350
00:20:11,482 --> 00:20:14,664
It's not just Okay, that's,
really interesting. I was

351
00:20:14,664 --> 00:20:18,386
assuming you were just, I, what
I had assumed is that you just

352
00:20:18,386 --> 00:20:22,048
have the sheet and then you have
a model that says we applied

353
00:20:22,048 --> 00:20:25,590
this to this. force this point
force at these locations and

354
00:20:25,590 --> 00:20:29,432
then we scanned it and this was
the result and that would inform

355
00:20:29,432 --> 00:20:33,274
the model but going down to like
current on motors on the arm is

356
00:20:33,274 --> 00:20:36,756
really awesome because i guess
they're non linear in their

357
00:20:36,756 --> 00:20:40,599
response most likely and so then
modeling that actually probably

358
00:20:40,599 --> 00:20:43,000
makes a lot of sense Sounds hard
though.

359
00:20:43,000 --> 00:20:46,730
Edward Mehr: yeah, so the robot
itself, and you're in robotics,

360
00:20:46,730 --> 00:20:50,579
so you fully appreciate this, is
depending on the pose, you might

361
00:20:50,579 --> 00:20:53,836
get different, deflection
profiles, right? If a robot's

362
00:20:53,836 --> 00:20:57,329
completely, extended out. It
might deflect much easier with

363
00:20:57,329 --> 00:21:00,645
less force. So the K constant
for it is much, much, more

364
00:21:00,645 --> 00:21:04,316
flexible than like when it is
in, in a very stiff position. So

365
00:21:04,316 --> 00:21:07,691
we need to actually can take
into account both the system

366
00:21:07,691 --> 00:21:11,303
itself, how it's responding to
the process, and then also the

367
00:21:11,303 --> 00:21:14,974
material, deforming, what forces
you're required to deform the

368
00:21:14,974 --> 00:21:18,289
sheet. But what kind of
information you're going to get,

369
00:21:18,289 --> 00:21:22,020
what kind of springback you want
to get. So you want to capture

370
00:21:22,020 --> 00:21:25,395
that at every step of the
process. And the more data, the

371
00:21:25,395 --> 00:21:29,007
more granular you can capture
the data, the more easier it is

372
00:21:29,007 --> 00:21:32,678
for you to make the model. But
more importantly, the more data

373
00:21:32,678 --> 00:21:36,171
you have to build a model,
because a lot of these processes

374
00:21:36,171 --> 00:21:39,901
require a lot of data. So if you
get had to start an end model,

375
00:21:39,901 --> 00:21:43,750
you have to make a lot of parts.
But if I could get deflection at

376
00:21:43,750 --> 00:21:46,000
every step. point of couple
hours that

377
00:21:46,000 --> 00:21:48,000
Audrow Nash: You get a lot more
data. Yeah

378
00:21:48,000 --> 00:21:49,000
Edward Mehr: I have a lot more
data.

379
00:21:49,000 --> 00:21:53,500
Audrow Nash: It's a lot easier
to converge to a good model with

380
00:21:53,500 --> 00:21:55,000
that I would imagine.

381
00:21:55,000 --> 00:21:56,000
Edward Mehr: Yes.

382
00:21:55,000 --> 00:21:59,377
Audrow Nash: One question just
for my understanding because I

383
00:21:59,377 --> 00:22:03,682
don't know too much about metal
forming You are, when you're

384
00:22:03,682 --> 00:22:08,202
metal forming you constrain all
the edges of the sheet of Metal

385
00:22:08,202 --> 00:22:11,000
that you're going to be forming,
right?

386
00:22:11,000 --> 00:22:12,000
Edward Mehr: Yes, you could. we
do

387
00:22:11,000 --> 00:22:14,000
Audrow Nash: it's sitting in a
picture frame almost, right?

388
00:22:14,000 --> 00:22:16,756
Edward Mehr: Yeah, you could do
that. that's what we do today.

389
00:22:16,756 --> 00:22:19,333
That's what, that's another
parameter. You could, modulate

390
00:22:19,333 --> 00:22:22,000
the amount of basic pressure
you're putting on the boundary.

391
00:22:22,000 --> 00:22:26,849
Audrow Nash: So then when you do
that, you're pushing into it and

392
00:22:26,849 --> 00:22:31,101
you're pinching part of the
metal out. Are you, so what I

393
00:22:31,101 --> 00:22:35,875
assume is that the metal gets a
bit thinner where it's stretched

394
00:22:35,875 --> 00:22:40,052
out. and so I would imagine
that's complex. I don't know

395
00:22:40,052 --> 00:22:44,827
between, so in a finished part,
how significant is the thinness,

396
00:22:44,827 --> 00:22:49,302
the thickness of the material?
Changed when you push it out,

397
00:22:49,302 --> 00:22:54,002
quite a bit. Does that, because
that would be another dimension

398
00:22:54,002 --> 00:22:58,776
in the modeling, I would imagine
the thickness of the metal part

399
00:22:58,776 --> 00:22:59,000
now

400
00:22:59,000 --> 00:23:03,561
Edward Mehr: Yes, so it depends
on what path you take to deform

401
00:23:03,561 --> 00:23:08,195
the material. So if I start from
a flat sheet and just, create a

402
00:23:08,195 --> 00:23:12,756
60 degree wall angle right on my
first layer as I'm forming the

403
00:23:12,756 --> 00:23:17,172
part, then the calculation is
relatively easy. It's basically

404
00:23:17,172 --> 00:23:21,371
a volume, yes, it's a volume
preservation, basically, law,

405
00:23:21,371 --> 00:23:25,932
right? Okay. I. Got, start from
a X millimeter thickness sheet.

406
00:23:25,932 --> 00:23:30,276
I deform it to 60 degree wall
angle. So it's going to reduce

407
00:23:30,276 --> 00:23:34,765
by cosine of 60 degrees times
the original thickness right? So

408
00:23:34,765 --> 00:23:39,181
simple, trigonometry will give
you the answer. But if I, but,

409
00:23:39,181 --> 00:23:43,742
if, for example, if I did first
a 45 degree wall angle and then

410
00:23:43,742 --> 00:23:47,000
push the 45 degree into 60
degree wall angle,

411
00:23:47,000 --> 00:23:48,000
Audrow Nash: Now it

412
00:23:47,000 --> 00:23:50,875
Edward Mehr: my thickness
profile is completely different,

413
00:23:50,875 --> 00:23:55,084
right? It's cosine of 45 times
cosine of 60, which is a smaller

414
00:23:55,084 --> 00:23:58,958
number. So actually you get
less. Less thinning, right? So

415
00:23:58,958 --> 00:24:03,100
another, so if you add thickness
preservation or set thickness

416
00:24:03,100 --> 00:24:07,242
targets for your geometry, your
final geometry, which you can,

417
00:24:07,242 --> 00:24:11,117
that also affects in what order
you want to form the part,

418
00:24:11,117 --> 00:24:15,259
right? You might want to form
intermediate geometries and then

419
00:24:15,259 --> 00:24:18,399
push those intermediate
geometries to the final

420
00:24:18,399 --> 00:24:19,000
geometry.

421
00:24:19,000 --> 00:24:24,019
Audrow Nash: Gotcha. Very
interesting. Okay. going back to

422
00:24:24,019 --> 00:24:29,644
the data that you're generating.
if you are observing kind of the

423
00:24:29,644 --> 00:24:34,750
robot state, also, robot, I
just, I'm, imagining your state

424
00:24:34,750 --> 00:24:39,942
space is blowing up and this
becomes a very complex equation

425
00:24:39,942 --> 00:24:44,962
because you use two, seven
degree of freedom robotic arms.

426
00:24:44,962 --> 00:24:46,000
Is it right?

427
00:24:46,000 --> 00:24:47,000
Edward Mehr: Correct.

428
00:24:47,000 --> 00:24:52,000
Audrow Nash: Okay. And you're
observing the, current at each

429
00:24:52,000 --> 00:24:56,000
joint in the seven degree of
freedom robot arms.

430
00:24:56,000 --> 00:24:59,846
Edward Mehr: Yeah, you could,
you have access to the current

431
00:24:59,846 --> 00:25:01,000
at each edge. Yes.

432
00:25:01,000 --> 00:25:04,969
Audrow Nash: I'm imagining it's
a very challenging optimization

433
00:25:04,969 --> 00:25:08,181
because it's very high
dimension. And maybe you use

434
00:25:08,181 --> 00:25:12,087
something like neural networks
that are very expressive to fit

435
00:25:12,087 --> 00:25:15,929
this. but then you might get a
big risk of overfitting and it

436
00:25:15,929 --> 00:25:19,520
might be tough to generalize.
just, tell me about some of

437
00:25:19,520 --> 00:25:23,488
those challenges of working in a
high dimensional space and how

438
00:25:23,488 --> 00:25:25,000
you guys deal with this.

439
00:25:25,000 --> 00:25:27,650
Edward Mehr: Yeah, so when
you're in high dimensional

440
00:25:27,650 --> 00:25:30,850
space, you have two challenges.
basically you have two routes of

441
00:25:30,850 --> 00:25:34,000
making models that are accurate.
One is generate a lot of data,

442
00:25:34,000 --> 00:25:36,000
Audrow Nash: Which you're doing.

443
00:25:36,000 --> 00:25:39,179
Edward Mehr: which, is something
you can do and we can, we are

444
00:25:39,179 --> 00:25:42,359
doing as well. Or in the short
term. in the earlier days where

445
00:25:42,359 --> 00:25:45,538
you don't have enough data, you
can also try to break down the

446
00:25:45,538 --> 00:25:48,000
problem, basically do physics
informed modeling.

447
00:25:48,000 --> 00:25:50,000
Audrow Nash: Yeah,

448
00:25:50,000 --> 00:25:51,000
Edward Mehr: So you can start

449
00:25:50,000 --> 00:25:52,000
Audrow Nash: you reduce your
space for

450
00:25:52,000 --> 00:25:54,000
Edward Mehr: introduce your
space.

451
00:25:54,000 --> 00:25:57,324
Audrow Nash: by leveraging
things like physics, simulation.

452
00:25:57,324 --> 00:25:58,000
Okay, I see.

453
00:25:58,000 --> 00:26:01,927
Edward Mehr: For example, you
can say, okay, if my robot is

454
00:26:01,927 --> 00:26:06,121
deflecting, I can either create
a deflection model based on the

455
00:26:06,121 --> 00:26:09,981
currents and the joint angles,
right? and that would be my

456
00:26:09,981 --> 00:26:14,175
space. Or I can just say most of
the deflection happens in each

457
00:26:14,175 --> 00:26:18,368
joint and, then model each joint
differently and then calculate

458
00:26:18,368 --> 00:26:22,495
the total deflection based on
the configuration of the joints.

459
00:26:22,495 --> 00:26:26,423
So now you're taking advantage
of kinematics and physics to

460
00:26:26,423 --> 00:26:30,416
say, okay, I can simplify. I
almost feature engineering, and

461
00:26:30,416 --> 00:26:34,543
they, call it in the data world,
like I am informing it. Maybe

462
00:26:34,543 --> 00:26:38,670
there's a better way I can solve
half of it for you as long as

463
00:26:38,670 --> 00:26:41,000
you tell me how each joint
deflect.

464
00:26:41,000 --> 00:26:42,000
Audrow Nash: Yeah,

465
00:26:42,000 --> 00:26:43,743
Edward Mehr: and then, so those
are some of the things we did

466
00:26:43,743 --> 00:26:44,000
early on.

467
00:26:44,000 --> 00:26:47,019
Audrow Nash: So you're
compressing your dimensions,

468
00:26:47,019 --> 00:26:50,215
basically, so you have fewer
dimensions, and it's more

469
00:26:50,215 --> 00:26:53,885
tractable with whatever data you
have. How do you go from zero

470
00:26:53,885 --> 00:26:57,614
data to some data? Do you start
with physics based and you just

471
00:26:57,614 --> 00:27:01,047
poke it a few times and see what
happens and then start to

472
00:27:01,047 --> 00:27:03,000
bootstrap a model or how is
that?

473
00:27:03,000 --> 00:27:06,389
Edward Mehr: That's exactly what
we did, right? Do like back of

474
00:27:06,389 --> 00:27:09,509
the envelope, not even physics
space, back of the envelope

475
00:27:09,509 --> 00:27:12,522
calculation initially. I
remember our initial deflection

476
00:27:12,522 --> 00:27:15,535
compensation model wasn't even
in joint space. It was in

477
00:27:15,535 --> 00:27:18,547
Cartesian space, which has
nothing to do with the robot,

478
00:27:18,547 --> 00:27:20,000
but we're like, okay, maybe

479
00:27:20,000 --> 00:27:22,000
Audrow Nash: so I would have
approached it too.

480
00:27:22,000 --> 00:27:25,404
Edward Mehr: like they're like
different in z direction as

481
00:27:25,404 --> 00:27:28,749
opposed to x y, so I can do a
Cartesian based, deflection

482
00:27:28,749 --> 00:27:32,330
model. That took us a long way
until we start seeing tears in

483
00:27:32,330 --> 00:27:35,910
very complex parts, and then
we're like, okay, This is model,

484
00:27:35,910 --> 00:27:39,255
this is pose dependent, so we
need to start, increase the

485
00:27:39,255 --> 00:27:43,011
complexity of the model. But at
that point, you have enough more

486
00:27:43,011 --> 00:27:46,239
data, enough data to start
looking into what that model

487
00:27:46,239 --> 00:27:48,000
will look like. Yeah, the key,

488
00:27:48,000 --> 00:27:50,000
Audrow Nash: Oh, go ahead.

489
00:27:50,000 --> 00:27:54,067
Edward Mehr: I think the key for
basically, I was listening to

490
00:27:54,067 --> 00:27:57,412
this podcast with Andrej
Karpathy about why, or no,

491
00:27:57,412 --> 00:28:01,545
actually it had to be with Ilya,
about why, ChatGPT exists, but

492
00:28:01,545 --> 00:28:05,349
not, A lot, it's much harder to
do foundational models for

493
00:28:05,349 --> 00:28:09,416
robots. key is because you need
to gather capture data. And in

494
00:28:09,416 --> 00:28:13,221
order to capture data, you need
to operate a huge fleet of

495
00:28:13,221 --> 00:28:17,288
robots. Now, yes. And then this,
means that you need to figure

496
00:28:17,288 --> 00:28:21,552
out a way to find an application
where you can create value based

497
00:28:21,552 --> 00:28:25,618
on heuristic simple models to
begin with, but that application

498
00:28:25,618 --> 00:28:29,029
after the data can have a
potential to significantly

499
00:28:29,029 --> 00:28:33,096
improve. but you need to find
that application. That's a tough

500
00:28:33,096 --> 00:28:36,835
part. Find an application where
even today with heuristic

501
00:28:36,835 --> 00:28:39,000
models, you can provide
benefits.

502
00:28:39,000 --> 00:28:42,667
Audrow Nash: I think Electric
Sheep, who was the first

503
00:28:42,667 --> 00:28:46,809
interview on this podcast, is
doing a great job of this. They

504
00:28:46,809 --> 00:28:51,222
are doing lawn mowing. With very
small mowers. And I think that's

505
00:28:51,222 --> 00:28:55,432
a very strong application for
this low risk, simple, just like

506
00:28:55,432 --> 00:28:59,506
they're doing localization and
path planning and things. And

507
00:28:59,506 --> 00:29:03,852
they're learning a lot of that.
but I think there are probably a

508
00:29:03,852 --> 00:29:08,130
lot of other spaces that could
do a similar approach. So what I

509
00:29:08,130 --> 00:29:12,543
would suspect is that realizing,
okay, data, is really important.

510
00:29:12,543 --> 00:29:16,957
The more data you have, the more
expressive your model can be and

511
00:29:16,957 --> 00:29:21,370
the better it will perform. so I
imagine. That realizing that you

512
00:29:21,370 --> 00:29:25,580
decided to grab all the data you
could. and then you've slowly

513
00:29:25,580 --> 00:29:29,722
been increasing the number of
dimensions that you're actually

514
00:29:29,722 --> 00:29:33,864
using in your optimization. Is
that true? And then also, how,

515
00:29:33,864 --> 00:29:38,074
where are we now in terms of,
like, how much is in the models?

516
00:29:38,074 --> 00:29:42,012
How expressive are the models?
How many dimensions are you

517
00:29:42,012 --> 00:29:45,000
including in the models? This
kind of thing.

518
00:29:45,000 --> 00:29:48,471
Edward Mehr: So yes, you're
right. So early days, basically

519
00:29:48,471 --> 00:29:51,647
build a lot of stuff based on
heuristic or back of the

520
00:29:51,647 --> 00:29:53,000
envelope calculation or

521
00:29:53,000 --> 00:29:55,000
Audrow Nash: Heuristics still.

522
00:29:55,000 --> 00:29:59,330
Edward Mehr: Yeah. And, and then
we start forming parts, like I

523
00:29:59,330 --> 00:30:03,180
said, early days, we had 25
trials before we get a part,

524
00:30:03,180 --> 00:30:07,579
right. Now we're down to five or
six. And initial models we used

525
00:30:07,579 --> 00:30:11,222
a lot of modeling, the
techniques we use were so very

526
00:30:11,222 --> 00:30:15,208
simple. Like you could even
start from regression. because

527
00:30:15,208 --> 00:30:19,401
once you break it down into a
simple thing, you can think of,

528
00:30:19,401 --> 00:30:23,388
let's say, a joint deflection as
a simple, one, one degree

529
00:30:23,388 --> 00:30:26,000
polynomial, right? because it's
mostly

530
00:30:26,000 --> 00:30:28,000
Audrow Nash: simple models for
these things.

531
00:30:28,000 --> 00:30:32,669
Edward Mehr: right? So, we start
from there. and I think last

532
00:30:32,669 --> 00:30:37,338
year, so to give you a little
bit of concept, like last year,

533
00:30:37,338 --> 00:30:41,778
it's actually a paper that we,
published with Northwestern

534
00:30:41,778 --> 00:30:46,447
around predicting the forces in
our process using graph based

535
00:30:46,447 --> 00:30:51,346
neural networks. so that's like
the, state of the art last year.

536
00:30:51,346 --> 00:30:56,091
and it, starts allowing you to
do cool things. for example, in

537
00:30:56,091 --> 00:30:59,000
that, paper, we use transfer
learning.

538
00:30:59,000 --> 00:31:00,000
Audrow Nash: what? Transfer
learning?

539
00:31:00,000 --> 00:31:02,701
Edward Mehr: Transfer learning,
and the way it works is that

540
00:31:02,701 --> 00:31:05,582
instead of trying to make a lot
of assumptions about our system,

541
00:31:05,582 --> 00:31:07,878
as you can imagine, you
mentioned yourself, is that

542
00:31:07,878 --> 00:31:10,714
there are a lot of parameters.
What if your calibration is even

543
00:31:10,714 --> 00:31:13,595
wrong to begin with, from system
to system? How does that affect

544
00:31:13,595 --> 00:31:14,000
the whole

545
00:31:14,000 --> 00:31:17,100
Audrow Nash: Or you could lump
them into additional dimensions

546
00:31:17,100 --> 00:31:18,000
that you're doing,

547
00:31:18,000 --> 00:31:21,159
Edward Mehr: Exactly, you could
lump them in an additional

548
00:31:21,159 --> 00:31:24,372
dimension, but then you need
more data. So what we ended up

549
00:31:24,372 --> 00:31:27,749
doing to answer your question,
in the short term, what can you

550
00:31:27,749 --> 00:31:30,799
do to actually improve the
performance of the models? We

551
00:31:30,799 --> 00:31:33,958
started using transfer learning.
We said, okay, with every

552
00:31:33,958 --> 00:31:37,280
system, first form 10 layers,
and then retrain the model that

553
00:31:37,280 --> 00:31:40,384
you have on just that 10 layers
in that machine with that

554
00:31:40,384 --> 00:31:43,598
process parameters. And then
predict what's going to happen

555
00:31:43,598 --> 00:31:46,811
for the rest of the part and
optimize for it based on that.

556
00:31:46,811 --> 00:31:49,752
And then we got much better
results because now you're

557
00:31:49,752 --> 00:31:52,802
learning the extrinsic very
quickly. different, specific

558
00:31:52,802 --> 00:31:56,342
features that each machine might
have or each configuration might

559
00:31:56,342 --> 00:31:59,610
have and learn those in the
first few layers, apply it using

560
00:31:59,610 --> 00:32:02,877
transfer learning, and then,
you've got to be more accurate.

561
00:32:02,877 --> 00:32:06,418
So that was the state of the art
last year. Now this year, we are

562
00:32:06,418 --> 00:32:09,903
going back again at that bigger
prize. Can I go all the way from

563
00:32:09,903 --> 00:32:13,389
this is input geometry, what is
the final design if I go through

564
00:32:13,389 --> 00:32:16,820
a certain set of parameters? So
that's what we're working on. I

565
00:32:16,820 --> 00:32:20,197
think we have enough data now to
capture that. so, stay tuned.

566
00:32:20,197 --> 00:32:23,682
I'll let you know how that model
goes. And we're pretty open. We

567
00:32:23,682 --> 00:32:27,113
publish most of the stuff we, we
do. Like I said, that paper is

568
00:32:27,113 --> 00:32:30,000
going to be in NAMRAC and it's
going to be presented.

569
00:32:30,000 --> 00:32:31,000
Audrow Nash: Hell yeah.

570
00:32:31,000 --> 00:32:34,828
Edward Mehr: But that's, the,
state of the art in terms of,

571
00:32:34,828 --> 00:32:38,981
the modeling today. We're still
using neural networks, but we're

572
00:32:38,981 --> 00:32:43,004
trying to basically go with the
bigger scope of the, model and

573
00:32:43,004 --> 00:32:46,897
not do small feature engineering
or, do things that are like

574
00:32:46,897 --> 00:32:48,000
physics informed.

575
00:32:48,000 --> 00:32:51,737
Audrow Nash: Very cool. Yeah.
You're removing a lot of the

576
00:32:51,737 --> 00:32:55,796
heuristics and seeing if you can
learn them and if that will be

577
00:32:55,796 --> 00:32:59,146
better. And then transfer
learning is very clever, I

578
00:32:59,146 --> 00:33:02,947
think, because you can learn a
good bit and you can make it

579
00:33:02,947 --> 00:33:06,877
general, but then you can learn
the very fine details on each

580
00:33:06,877 --> 00:33:11,000
specific machine, or with a new
material, or that kind of thing.

581
00:33:11,000 --> 00:33:14,163
Edward Mehr: It's funny,
craftsmen do that too. So I used

582
00:33:14,163 --> 00:33:17,604
to do sheet shaping myself, and
we used to, I used to work, at

583
00:33:17,604 --> 00:33:21,045
the shop, in Pomoda, we would
do, panels for hot rods. And you

584
00:33:21,045 --> 00:33:24,541
do it under a, PowerHammer, you
form it. But a lot of times you

585
00:33:24,541 --> 00:33:28,038
get a material, you don't know
what it is. they say, okay, this

586
00:33:28,038 --> 00:33:31,534
is a mild steel, maybe. And then
you hammer it a little bit and

587
00:33:31,534 --> 00:33:34,808
you get a feel of it. You were
like, okay, this is how this

588
00:33:34,808 --> 00:33:38,416
material operates. And then you,
estimate what you need to do and

589
00:33:38,416 --> 00:33:41,968
how long it's going to take you.
and I think it's the same thing

590
00:33:41,968 --> 00:33:45,353
with transfer learning, get the
robot to form a slight little

591
00:33:45,353 --> 00:33:48,627
part or like few layers of the
first part, and then capture

592
00:33:48,627 --> 00:33:52,013
that and use that data to figure
out and form the rest of the

593
00:33:52,013 --> 00:33:54,843
part. And that's something that
humans do. I had a,

594
00:33:54,843 --> 00:33:58,284
blacksmithing teacher, you learn
all these little tricks where

595
00:33:58,284 --> 00:34:01,725
oh, if you go to a junkyard and
you want to buy good metal, we

596
00:34:01,725 --> 00:34:04,000
used to carry these hand pocket
grinders.

597
00:34:04,000 --> 00:34:06,000
Audrow Nash: Oh, wow.

598
00:34:06,000 --> 00:34:08,675
Edward Mehr: and then you grind
it and based on the amount of

599
00:34:08,675 --> 00:34:11,000
sparks it makes, you can know
how much carbon it has.

600
00:34:11,000 --> 00:34:12,000
Audrow Nash: Oh,

601
00:34:12,000 --> 00:34:14,862
Edward Mehr: and the more carbon
it has, it's usually, yeah, a

602
00:34:14,862 --> 00:34:17,815
stronger material. But this is,
these are all concepts, like you

603
00:34:17,815 --> 00:34:20,631
try the material and you see the
feedback and response of the

604
00:34:20,631 --> 00:34:23,354
material, and you know exactly
what you're dealing with. So

605
00:34:23,354 --> 00:34:26,123
this is what craftsmen used to
do, like mental patterns that

606
00:34:26,123 --> 00:34:29,123
created for a long time. I think
now with robotic and AI systems,

607
00:34:29,123 --> 00:34:31,985
we can actually replicate that,
and maybe it's a faster way to

608
00:34:31,985 --> 00:34:33,000
scale these processes.

609
00:34:33,000 --> 00:34:35,960
Audrow Nash: Yeah, it's super
cool. Hell yeah. Tell me a

610
00:34:35,960 --> 00:34:38,868
little bit. So you mentioned
you're publishing research

611
00:34:38,868 --> 00:34:42,198
papers, which is very cool. I'm
glad you're contributing to the

612
00:34:42,198 --> 00:34:45,000
collective knowledge of the
whole robotics community.

613
00:34:45,000 --> 00:34:49,862
Audrow Nash: tell me a bit about
your team and Where are you? I

614
00:34:49,862 --> 00:34:54,569
know you're in LA, but tell me
about, just about your company

615
00:34:54,569 --> 00:34:59,277
in general. how many people you
are? How large is your space?

616
00:34:59,277 --> 00:35:03,598
What kind of things are you
working on? is it still very

617
00:35:03,598 --> 00:35:08,383
early researchy days or are you
having customers and all these

618
00:35:08,383 --> 00:35:09,000
details?

619
00:35:09,000 --> 00:35:13,026
Edward Mehr: Yeah, for sure. so
we started in 2019. and, we

620
00:35:13,026 --> 00:35:16,575
started in a facility. It's also
a, defense contract

621
00:35:16,575 --> 00:35:19,987
manufacturer. That's our
landlord. We got the back

622
00:35:19,987 --> 00:35:24,218
portion of this facility with
actually a little bit of shares.

623
00:35:24,218 --> 00:35:27,766
We paid the landlord rent in
shares and they're very

624
00:35:27,766 --> 00:35:31,861
supportive. but since then we
can, we have, we can expand it

625
00:35:31,861 --> 00:35:35,751
in the same facility. We have
35, 000 square feet in this

626
00:35:35,751 --> 00:35:39,777
facility. we've deployed 11
robotic cells downstairs, so 22

627
00:35:39,777 --> 00:35:43,667
robots in total. and then we
just acquired another space.

628
00:35:43,667 --> 00:35:47,966
It's a 66, 000 square feet. It's
two miles away. So now we have

629
00:35:47,966 --> 00:35:52,265
around 100, 110, 000 square feet
in total. we are slowly moving

630
00:35:52,265 --> 00:35:56,428
into data space as well. but the
goal of that space is not so

631
00:35:56,428 --> 00:36:00,045
much house our robotic cells to
increase the, rate of

632
00:36:00,045 --> 00:36:01,000
manufacturing.

633
00:36:01,000 --> 00:36:02,000
Audrow Nash: More data and
everything else.

634
00:36:02,000 --> 00:36:05,846
Edward Mehr: yes. So we are,
hoping to get to a point where,

635
00:36:05,846 --> 00:36:09,115
we can manufacture a
manufacturing cell in a month.

636
00:36:09,115 --> 00:36:12,000
so, we can get 12 manufacturing
cells a year,

637
00:36:12,000 --> 00:36:14,000
Audrow Nash: It's so meta.
That's great.

638
00:36:14,000 --> 00:36:17,910
Edward Mehr: Yeah. So, that's,
the company's right now close to

639
00:36:17,910 --> 00:36:21,945
70 people. We're planning to get
to around 130, 140 people by the

640
00:36:21,945 --> 00:36:23,000
end of next year.

641
00:36:23,000 --> 00:36:24,000
Audrow Nash: Wow. Damn.

642
00:36:24,000 --> 00:36:27,871
Edward Mehr: so we're growing
rapidly. the cost, the product

643
00:36:27,871 --> 00:36:28,000
is

644
00:36:28,000 --> 00:36:31,588
Audrow Nash: by the end of next
year? So in the next, what is

645
00:36:31,588 --> 00:36:34,000
it? I don't know, 18 months or
something?

646
00:36:34,000 --> 00:36:37,953
Edward Mehr: year and something.
Yes. mid next year, I think

647
00:36:37,953 --> 00:36:41,512
that's, Yes. and yeah, so the
company, almost has been

648
00:36:41,512 --> 00:36:45,597
doubling the staff every year.
and, it's already in customer's

649
00:36:45,597 --> 00:36:49,353
hands. early days of, our
company was about making parts.

650
00:36:49,353 --> 00:36:53,043
one thing I learned from my
previous job was that do not

651
00:36:53,043 --> 00:36:57,128
lock yourself in the lab. And by
the time you're done building

652
00:36:57,128 --> 00:37:00,818
it, the product is already
obsolete. so we ended up even

653
00:37:00,818 --> 00:37:04,442
the first time, and even we
didn't have robots, we were

654
00:37:04,442 --> 00:37:08,659
lucky enough to get customers to
be like, okay, what are some of

655
00:37:08,659 --> 00:37:12,744
the complex challenges you have
on sheet metal? Send it to us.

656
00:37:12,744 --> 00:37:16,764
We're going to, we're going to
work on those. so I think NASA

657
00:37:16,764 --> 00:37:20,783
and Air Force were one of our
early customers who provided us

658
00:37:20,783 --> 00:37:24,934
with the challenges they had. so
since day one, we had Customer

659
00:37:24,934 --> 00:37:28,756
Paying Parts, even our first
year of operation, I think we

660
00:37:28,756 --> 00:37:32,578
made 300, 000 with me and my co
founder, so it was, a good

661
00:37:32,578 --> 00:37:36,795
amount of interaction very early
on, and we've been able to like

662
00:37:36,795 --> 00:37:40,550
almost double the revenue or
more than double the revenue

663
00:37:40,550 --> 00:37:42,000
every year since then,

664
00:37:42,000 --> 00:37:44,000
Audrow Nash: Hell yeah.

665
00:37:44,000 --> 00:37:47,720
Edward Mehr: the system right
now, we do manufacture parts in

666
00:37:47,720 --> 00:37:51,561
our facility, but that's usually
in the goal of eventually them

667
00:37:51,561 --> 00:37:54,000
deploying the system in their,
in their,

668
00:37:54,000 --> 00:37:57,000
Audrow Nash: Do you want to sell
work cells effectively? Oh,

669
00:37:57,000 --> 00:38:00,937
Edward Mehr: Yeah, we already
have deployed ourselves into the

670
00:38:00,937 --> 00:38:04,238
customer facility. So for
example, one of them is at

671
00:38:04,238 --> 00:38:08,111
Warner Robins Air Force Base,
working on building, components

672
00:38:08,111 --> 00:38:11,540
for, sustainment and repair of
aircrafts. they have an

673
00:38:11,540 --> 00:38:15,413
interesting challenge in the
military is, they have, hundreds

674
00:38:15,413 --> 00:38:19,286
of weapon systems and defense
systems that are, some of them,

675
00:38:19,286 --> 00:38:23,159
70, 80 years old. think about B
52. and these are sheet metal

676
00:38:23,159 --> 00:38:27,286
airplanes, right? And every time
a component, let's say a landing

677
00:38:27,286 --> 00:38:31,095
gear door gets damaged, they
either have an inventory of it,

678
00:38:31,095 --> 00:38:34,778
but in a lot of cases they
don't, because these are like a

679
00:38:34,778 --> 00:38:38,587
70 year old plane. So they need
to have the dies for it, the

680
00:38:38,587 --> 00:38:42,651
tooling for it, and the tooling
also sometimes doesn't exist. So

681
00:38:42,651 --> 00:38:46,397
for example, for a landing gear
door of a certain aircraft,

682
00:38:46,397 --> 00:38:50,206
we're looking at four year lead
time and millions of dollars

683
00:38:50,206 --> 00:38:54,333
before you can get your part. so
the aircraft needs to be down on

684
00:38:54,333 --> 00:38:58,333
the ground for four years. it's
a huge, hugely affect the fleet

685
00:38:58,333 --> 00:39:02,333
readiness, for our military. so
we're working with them turning

686
00:39:02,333 --> 00:39:05,000
some of those four yearly times
into days.

687
00:39:05,000 --> 00:39:06,000
Audrow Nash: Hell yeah.

688
00:39:06,000 --> 00:39:08,836
Edward Mehr: Yeah, so that's
like one of our biggest

689
00:39:08,836 --> 00:39:12,164
customers. it's deployed there
and we're working with some of

690
00:39:12,164 --> 00:39:15,000
the other folks in automotive
and aerospace as well.

691
00:39:15,000 --> 00:39:19,429
Audrow Nash: Okay. 70 people,
two places in LA, 22 robots. How

692
00:39:19,429 --> 00:39:23,714
are you guys funded? So you
mentioned you're already getting

693
00:39:23,714 --> 00:39:28,000
revenue. but I assume you've
taken some investment rounds or

694
00:39:28,000 --> 00:39:32,429
how, have you guys been funded
up to this point other than the

695
00:39:32,429 --> 00:39:34,000
revenue you're making?

696
00:39:34,000 --> 00:39:38,447
Edward Mehr: Yeah, so we're a
venture backed company, which

697
00:39:38,447 --> 00:39:43,196
has its own challenges and good,
things and bad things. to this

698
00:39:43,196 --> 00:39:47,342
date, we've raised about 45
million in venture funding,

699
00:39:47,342 --> 00:39:49,000
right? So we're, very,

700
00:39:49,000 --> 00:39:50,000
Audrow Nash: What round, have
you

701
00:39:50,000 --> 00:39:53,000
Edward Mehr: We just raised our
Series B last year

702
00:39:53,000 --> 00:39:54,000
Audrow Nash: Oh, congrats. Oh
yeah,

703
00:39:54,000 --> 00:39:56,000
Edward Mehr: last year. Yeah.

704
00:39:56,000 --> 00:39:59,444
Audrow Nash: That's fantastic
because I hear that it gets much

705
00:39:59,444 --> 00:40:02,722
harder, like B and C are the two
hardest funding rounds for

706
00:40:02,722 --> 00:40:05,000
robotics, I think for hardware
in general

707
00:40:05,000 --> 00:40:08,694
Edward Mehr: yes. I think that
especially last year was pretty

708
00:40:08,694 --> 00:40:12,030
tough because, once you get to B
and C, you need to have

709
00:40:12,030 --> 00:40:15,902
commercial traction. you need to
show you have a good revenue and

710
00:40:15,902 --> 00:40:19,715
you also have good margins that
is suitable for a venture funded

711
00:40:19,715 --> 00:40:22,932
business. That's the one
challenge with venture funded

712
00:40:22,932 --> 00:40:26,626
business is that what is their
alternative? The alternative is

713
00:40:26,626 --> 00:40:30,021
a software or ChatGPT that it's
I don't know, what is the

714
00:40:30,021 --> 00:40:33,000
margins there? Like 80, 90%,
whatever it is today,

715
00:40:33,000 --> 00:40:34,000
Audrow Nash: I don't know.

716
00:40:33,000 --> 00:40:36,732
Edward Mehr: For us, so we need
to be able to compete with that.

717
00:40:36,732 --> 00:40:39,881
and it's much tougher for
robotic companies with CapEx

718
00:40:39,881 --> 00:40:43,497
expenditure to get there. Out.
But I think we have been making

719
00:40:43,497 --> 00:40:46,996
a lot of interesting choices and
hopefully wise choices that

720
00:40:46,996 --> 00:40:50,670
allows us to our margin to stay
up so we can compete with those

721
00:40:50,670 --> 00:40:53,877
type of, those types of
businesses, but I guess tougher

722
00:40:53,877 --> 00:40:57,609
for, as you pointed out in B and
C because economics is the main

723
00:40:57,609 --> 00:41:00,000
driver of raising a C round or a
B round.

724
00:41:00,000 --> 00:41:04,482
Audrow Nash: there are a lot of
good, I don't know what the word

725
00:41:04,482 --> 00:41:08,895
is, like headwinds or forces or
something. like we're reshoring

726
00:41:08,895 --> 00:41:13,167
a lot of manufacturing is what
it seems. and so I would think

727
00:41:13,167 --> 00:41:16,949
our government is very
supportive of what you guys are

728
00:41:16,949 --> 00:41:20,591
doing, which I think is
wonderful. And then, I don't

729
00:41:20,591 --> 00:41:25,004
know there's just probably a big
interest like just seeing what

730
00:41:25,004 --> 00:41:28,926
feels like the zeitgeist on X
it's like manufacturing is

731
00:41:28,926 --> 00:41:32,988
becoming cool wasn't the case
years ago and I think it's a

732
00:41:32,988 --> 00:41:36,000
wonderful thing for our country
in general.

733
00:41:36,000 --> 00:41:39,626
Edward Mehr: I agree. I agree. I
think this year, especially, I

734
00:41:39,626 --> 00:41:42,906
think the, leaf is turning. I
think last year was still a

735
00:41:42,906 --> 00:41:46,475
little bit tougher. Everybody
was in holding mode. I think now

736
00:41:46,475 --> 00:41:49,928
there's a lot of excitement
about manufacturing, but I think

737
00:41:49,928 --> 00:41:52,000
also at the same time that puts
the,

738
00:41:52,000 --> 00:41:53,000
Audrow Nash: High expectations

739
00:41:53,000 --> 00:41:56,542
Edward Mehr: Expectations on it,
which is in a good way under

740
00:41:56,542 --> 00:41:59,561
robotics and hardware and
manufacturing companies to

741
00:41:59,561 --> 00:42:02,000
really go about this in a smart
way. Yeah.

742
00:42:02,000 --> 00:42:05,059
Audrow Nash: Especially with
cash being so expensive

743
00:42:05,059 --> 00:42:08,824
relatively now because then yeah
you have to be pragmatic or you

744
00:42:08,824 --> 00:42:11,000
go out of business this kind of
thing

745
00:42:11,000 --> 00:42:13,919
Edward Mehr: Yes. So turn this
excitement and make it

746
00:42:13,919 --> 00:42:17,335
sustainable, right? So, that all
these companies that come in,

747
00:42:17,335 --> 00:42:20,695
obviously a lot of them will be
successful, some won't. but I

748
00:42:20,695 --> 00:42:23,945
think it's new grounds. we have
to think about our business

749
00:42:23,945 --> 00:42:27,250
models are not going to be the
same as SaaS business models.

750
00:42:27,250 --> 00:42:30,280
Our, the way we're going to make
margins is going to be

751
00:42:30,280 --> 00:42:33,199
different, but we need to pave
those paths. Like it's

752
00:42:33,199 --> 00:42:36,504
interesting that even investors.
Don't know exactly what the

753
00:42:36,504 --> 00:42:40,085
model will look like, but if the
margins are not 60 70 percent at

754
00:42:40,085 --> 00:42:43,390
the end of the day, it's tougher
to compete with some of the

755
00:42:43,390 --> 00:42:46,530
alternatives they have, so we
need to be very, we need to

756
00:42:46,530 --> 00:42:50,000
think about it extra hard and be
very adaptable and agile here.

757
00:42:50,000 --> 00:42:53,000
Audrow Nash: Yeah, so what was
your valuation?

758
00:42:53,000 --> 00:42:54,000
Edward Mehr: Yeah,

759
00:42:54,000 --> 00:42:56,000
Audrow Nash: I don't know if
that's public, but

760
00:42:56,000 --> 00:43:02,000
Edward Mehr: yeah, It's not
public, but we raised relatively

761
00:43:02,000 --> 00:43:03,000
with good,

762
00:43:03,000 --> 00:43:05,000
Audrow Nash: Didn't trade that
much equity for everything.

763
00:43:05,000 --> 00:43:08,881
Edward Mehr: Yeah, so we, have
been every round, we at least

764
00:43:08,881 --> 00:43:12,373
doubled or tripled our, more
than double and sometimes

765
00:43:12,373 --> 00:43:16,318
tripled our valuation, from the
beginning. I won't be able to

766
00:43:16,318 --> 00:43:18,000
give you the exact amount,

767
00:43:18,000 --> 00:43:19,000
Audrow Nash: Yeah, ballpark
would be

768
00:43:19,000 --> 00:43:20,000
Edward Mehr: it has been
increased. Again, we

769
00:43:20,000 --> 00:43:25,276
Audrow Nash: has been increased.
It was a good round, basically.

770
00:43:25,276 --> 00:43:30,222
Good, yeah, because a lot of
companies have been seeing down

771
00:43:30,222 --> 00:43:35,086
rounds, I think, which just, I
don't know, cash is becoming

772
00:43:35,086 --> 00:43:40,032
tighter, or less. I just, had an
interview, With a person in

773
00:43:40,032 --> 00:43:45,143
venture capital and they were
saying that in 2021-2022 kind of

774
00:43:45,143 --> 00:43:49,595
thing, it was just like everyone
was throwing money at

775
00:43:49,595 --> 00:43:54,953
everything. And so people raised
huge up rounds and then, now the

776
00:43:54,953 --> 00:43:59,735
expectations are coming more
back down to earth. And then,

777
00:43:59,735 --> 00:44:04,928
That has created some friction
for a lot of companies and a lot

778
00:44:04,928 --> 00:44:06,000
of investors.

779
00:44:06,000 --> 00:44:09,218
Edward Mehr: Yeah, no, you're
absolutely right. I think, we

780
00:44:09,218 --> 00:44:12,000
were lucky that every round is
more than double the

781
00:44:12,000 --> 00:44:15,000
Audrow Nash: Yep. So you guys
are doing great. Hell yeah.

782
00:44:15,000 --> 00:44:18,504
Edward Mehr: I think part of
that is as tempting as it is to

783
00:44:18,504 --> 00:44:21,949
get a very high valuation in
early days, I think it is very

784
00:44:21,949 --> 00:44:25,628
important to, almost pace your
progress and set milestones in a

785
00:44:25,628 --> 00:44:29,307
way where you don't set yourself
up for a down round. I think a

786
00:44:29,307 --> 00:44:32,869
lot of folks come in and, tell
very amazing stories, which is

787
00:44:32,869 --> 00:44:36,489
great. I'll get the biggest one
I can, and then it's okay, the

788
00:44:36,489 --> 00:44:39,759
milestone now I'm setting up
because if the valuation is

789
00:44:39,759 --> 00:44:43,263
high, the expectation is high
too. So the milestones you set

790
00:44:43,263 --> 00:44:46,766
up for yourself for the next
round, if you don't meet those,

791
00:44:46,766 --> 00:44:50,036
then You know, the valuation
drops and that's not a good

792
00:44:50,036 --> 00:44:53,657
thing for your employees that
are stakeholders in the company.

793
00:44:53,657 --> 00:44:57,336
Obviously not a good thing for
your investors as well, but more

794
00:44:57,336 --> 00:45:00,489
importantly for the employees
who are, leaving jobs at

795
00:45:00,489 --> 00:45:03,934
Microsoft and Google to join
your company. I want to always

796
00:45:03,934 --> 00:45:07,438
show them that, this is a better
alternative than staying in

797
00:45:07,438 --> 00:45:10,650
those places. If you have a
downturn, that's not a good

798
00:45:10,650 --> 00:45:11,000
thing.

799
00:45:11,000 --> 00:45:16,753
Audrow Nash: How, so another
thing that's interesting to me,

800
00:45:16,753 --> 00:45:22,507
you guys are in LA. Tell me
about the LA, and actually, how,

801
00:45:22,507 --> 00:45:28,260
how close to, the major part of
LA are you, far out east, or

802
00:45:28,260 --> 00:45:32,000
where, where, are you in the LA
region?

803
00:45:32,000 --> 00:45:34,432
Edward Mehr: Yeah, if you're
familiar with it, I think we're

804
00:45:34,432 --> 00:45:35,000
in the Valley.

805
00:45:35,000 --> 00:45:36,000
Audrow Nash: Oh, hell yeah. So

806
00:45:36,000 --> 00:45:40,239
Edward Mehr: it is Los Angeles,
the city. and, you're probably

807
00:45:40,239 --> 00:45:44,000
half an hour away from downtown,
if there's no traffic.

808
00:45:44,000 --> 00:45:46,000
Audrow Nash: If there's no
traffic. Yeah.

809
00:45:46,000 --> 00:45:49,643
Edward Mehr: Yes. for like maybe
45 minutes away from SpaceX, for

810
00:45:49,643 --> 00:45:52,501
those people who are like
anchoring a lot of harbor

811
00:45:52,501 --> 00:45:55,751
companies around the SpaceX in
South Bay. So we are on the

812
00:45:55,751 --> 00:45:59,170
other side of the LA. So South
Bay is traditionally have been

813
00:45:59,170 --> 00:46:02,757
aerospace manufacturing hub. We
are in the other matter of fact,

814
00:46:02,757 --> 00:46:06,231
in the other manufacturing hub,
Chatsworth, which is North end

815
00:46:06,231 --> 00:46:09,650
of the LA area in the Valley. We
have companies like a lot of

816
00:46:09,650 --> 00:46:13,068
aerospace companies and machine
shops here as well. So it's a

817
00:46:13,068 --> 00:46:16,655
big manufacturing hub. it's one
of the other, basically it's the

818
00:46:16,655 --> 00:46:18,000
other manufacturing hub,

819
00:46:18,000 --> 00:46:21,000
Audrow Nash: What's the other
one from that?

820
00:46:21,000 --> 00:46:22,000
Edward Mehr: South Bay, like

821
00:46:22,000 --> 00:46:23,000
Audrow Nash: Oh, other one in
LA. I see.

822
00:46:23,000 --> 00:46:27,148
Edward Mehr: yeah, El Segundo,
Torrance, Hawthorne, that

823
00:46:27,148 --> 00:46:31,000
region, which is like 45 minutes
away from us, yeah.

824
00:46:31,000 --> 00:46:35,370
Audrow Nash: And so do you think
it's a very good thing for you

825
00:46:35,370 --> 00:46:39,671
guys to be very close to like
relatively close to SpaceX and a

826
00:46:39,671 --> 00:46:43,000
lot of military bases? is it a
really good spot?

827
00:46:43,000 --> 00:46:47,726
Edward Mehr: I think so. I think
we have a good ability to

828
00:46:47,726 --> 00:46:52,126
attract hardware talent here.
Software is a little bit

829
00:46:52,126 --> 00:46:54,000
tougher, right? Because

830
00:46:54,000 --> 00:46:56,000
Audrow Nash: Oh, they're in SF
or in the

831
00:46:56,000 --> 00:47:01,000
Edward Mehr: Area, or maybe East
Coast a little bit, Boston,

832
00:47:01,000 --> 00:47:02,000
Audrow Nash: Oh,

833
00:47:02,000 --> 00:47:05,796
Edward Mehr: and New York. So
it's a little bit tougher on the

834
00:47:05,796 --> 00:47:09,408
software side. But overall, I
think, the hardware talent is

835
00:47:09,408 --> 00:47:11,000
much more accessible here.

836
00:47:11,000 --> 00:47:15,701
Audrow Nash: Gotcha. Very cool.
Let's see. So pivoting a little

837
00:47:15,701 --> 00:47:16,000
bit,

838
00:47:16,000 --> 00:47:23,134
Audrow Nash: why haven't we seen
many big and successful robotics

839
00:47:23,134 --> 00:47:25,000
companies yet? So

840
00:47:25,000 --> 00:47:29,162
Edward Mehr: Yeah, I think
nobody really has figured out

841
00:47:29,162 --> 00:47:33,919
what is the right business model
for robotics companies. I think

842
00:47:33,919 --> 00:47:38,230
for a robotic company to be
successful, for any company to

843
00:47:38,230 --> 00:47:42,243
be successful, I think,
especially a startup, you need

844
00:47:42,243 --> 00:47:46,257
to almost provide 10x
improvements. the reason for 10x

845
00:47:46,257 --> 00:47:50,270
improvement is that, as a
startup, most, you're mostly

846
00:47:50,270 --> 00:47:54,878
odds are against your success.
You might run out of cash, your

847
00:47:54,878 --> 00:47:59,338
talent might not come in, you
might make a few mistakes that

848
00:47:59,338 --> 00:48:03,946
causes your company to die. So
you're really prone to failure.

849
00:48:03,946 --> 00:48:08,331
So if the opportunity you're
going after is not 10x better,

850
00:48:08,331 --> 00:48:12,939
the outcome is going to be 10
times better, at least, than the

851
00:48:12,939 --> 00:48:17,399
current outcome. Then you're,
you have very limited room for

852
00:48:17,399 --> 00:48:20,000
mistakes because if you're 10x,
the

853
00:48:20,000 --> 00:48:23,000
Audrow Nash: So you view it as
margin in a sense,

854
00:48:23,000 --> 00:48:24,000
Edward Mehr: in

855
00:48:24,000 --> 00:48:26,672
Audrow Nash: the amount of times
better. It's yeah. So if you,

856
00:48:26,672 --> 00:48:29,000
only do two times better, you
have a very small margin

857
00:48:29,000 --> 00:48:32,108
Edward Mehr: yeah, you make few
mistakes and now you're as good

858
00:48:32,108 --> 00:48:35,018
as the traditional technology,
right? And then you're like,

859
00:48:35,018 --> 00:48:38,175
okay, is it really better? Do I
want to pay more for robotics in

860
00:48:38,175 --> 00:48:41,381
this space? and I think that's a
problem. A lot of companies had,

861
00:48:41,381 --> 00:48:44,390
there are a lot of incremental
improvements and I don't think

862
00:48:44,390 --> 00:48:47,054
you can build a company on
incremental improvement, at

863
00:48:47,054 --> 00:48:50,161
least in this space. In finance,
maybe you can, but in business

864
00:48:50,161 --> 00:48:51,000
space, you can't,

865
00:48:51,000 --> 00:48:53,417
Audrow Nash: And this space
being robotics, you're saying,

866
00:48:53,417 --> 00:48:54,000
or technology?

867
00:48:54,000 --> 00:48:56,000
Edward Mehr: and slash hardware,
I would say,

868
00:48:56,000 --> 00:48:58,000
Audrow Nash: Slash hardware. So

869
00:48:58,000 --> 00:49:02,100
Edward Mehr: need to have that
room for, failures. And say,

870
00:49:02,100 --> 00:49:06,479
okay, despite all the failures,
maybe I end up being 2x better.

871
00:49:06,479 --> 00:49:10,440
I started at 10x and then I
ended up being 2x better, but

872
00:49:10,440 --> 00:49:14,888
still 2x better, right? so, that
has been, I think, a story with

873
00:49:14,888 --> 00:49:16,000
a lot of robotic

874
00:49:16,000 --> 00:49:19,471
Audrow Nash: a really strong
concept initially that creates

875
00:49:19,471 --> 00:49:21,000
really great improvements.

876
00:49:21,000 --> 00:49:24,416
Edward Mehr: You almost have to
be very ambitious initially and

877
00:49:24,416 --> 00:49:27,614
have a path to make a very
ambitious company work. And then

878
00:49:27,614 --> 00:49:30,759
while you're shooting for the
stars, you maybe land on the

879
00:49:30,759 --> 00:49:33,904
moon, right? But, but that's, I
think that definitely much

880
00:49:33,904 --> 00:49:37,102
definitely applies in this
space. And that's why we haven't

881
00:49:37,102 --> 00:49:39,000
seen a lot of successful
companies.

882
00:49:39,000 --> 00:49:43,903
Audrow Nash: Interesting. What
are some examples of, naming

883
00:49:43,903 --> 00:49:47,974
specific ones that haven't
innovated, innovative,

884
00:49:47,974 --> 00:49:53,126
innovated, 10 times, possibly
not be good, but can you give me

885
00:49:53,126 --> 00:49:58,112
some like cartoons of companies
that have not been ambitious

886
00:49:58,112 --> 00:50:03,097
enough, and then I'd love to
hear ones from your perspective

887
00:50:03,097 --> 00:50:08,000
that are. Really pushing it, 10
times or more to be better.

888
00:50:08,000 --> 00:50:11,313
Edward Mehr: it's a good
question. the example I give

889
00:50:11,313 --> 00:50:15,188
usually is, to some extent, I
come from 3D printing world, and

890
00:50:15,188 --> 00:50:19,063
I think that could potentially
be, if you consider 3D printing

891
00:50:19,063 --> 00:50:22,000
machines as robotic machines,
they're automated

892
00:50:22,000 --> 00:50:24,000
Audrow Nash: Yeah, totally. They
sense, they think, they act.

893
00:50:24,000 --> 00:50:27,862
Edward Mehr: I think that is one
area where, we initially thought

894
00:50:27,862 --> 00:50:31,071
it was like going to be
revolutionary, but I think the

895
00:50:31,071 --> 00:50:34,754
reach of the parts it can do is
not as big as we thought. that

896
00:50:34,754 --> 00:50:38,320
was the point. And if it was
providing advantages in some of

897
00:50:38,320 --> 00:50:41,766
the applications, it was not 10
times better. Like in some

898
00:50:41,766 --> 00:50:45,271
applications, it was like, if it
was like you're building a

899
00:50:45,271 --> 00:50:48,836
rocket engine, a heat exchanger
additive is really good. And

900
00:50:48,836 --> 00:50:52,461
what if you're making a bracket
for some vehicle? Yeah. argue

901
00:50:52,461 --> 00:50:56,323
your way one way or another, but
it's not 10 times better. So you

902
00:50:56,323 --> 00:51:00,126
invest a lot of money. There's a
lot of room for failure. And at

903
00:51:00,126 --> 00:51:03,809
the end of the day, it's not 10
times better. It doesn't drive

904
00:51:03,809 --> 00:51:07,315
Adoption. So I think that was
one example and that's what I

905
00:51:07,315 --> 00:51:10,939
learned actually from, when I
was in 3D printing, look at the

906
00:51:10,939 --> 00:51:14,623
market size of the 3D printing
right now is around 16 billion,

907
00:51:14,623 --> 00:51:17,000
everything, plastics, metal,
everything,

908
00:51:17,000 --> 00:51:19,000
Audrow Nash: Yeah.

909
00:51:19,000 --> 00:51:22,069
Edward Mehr: and there's
probably close to like somewhere

910
00:51:22,069 --> 00:51:25,408
from 6 to 10 billion dollars in
investment that went into that

911
00:51:25,408 --> 00:51:26,000
technology.

912
00:51:26,000 --> 00:51:27,000
Audrow Nash: Not that much of a
return.

913
00:51:27,000 --> 00:51:29,000
Edward Mehr: So not, much of a
return there, right?

914
00:51:29,000 --> 00:51:30,884
Audrow Nash: And there's a whole
bunch of failures and everything

915
00:51:30,884 --> 00:51:31,000
too.

916
00:51:31,000 --> 00:51:34,978
Edward Mehr: There's a lot of
failures. Yep. yeah. Now, when

917
00:51:34,978 --> 00:51:39,156
we started this technology, for
example, like one of the things

918
00:51:39,156 --> 00:51:42,670
I really wanted to go after a
market segment that has

919
00:51:42,670 --> 00:51:46,914
significant room. so that's why
we went after sheet forming as a

920
00:51:46,914 --> 00:51:50,959
first, foray into manufacturing
in this company. and that's a

921
00:51:50,959 --> 00:51:55,004
250 billion market. so lots of
room for improvement, but then

922
00:51:55,004 --> 00:51:58,916
also, started thinking about,
okay, what is the portion I'm

923
00:51:58,916 --> 00:52:02,695
replacing? I'm replacing the
dyes, and I want to go after

924
00:52:02,695 --> 00:52:06,674
sector that the dyes are very,
expensive, right? or the lead

925
00:52:06,674 --> 00:52:08,000
times are very high.

926
00:52:08,000 --> 00:52:11,000
Audrow Nash: Military was a
great thing for that then.

927
00:52:11,000 --> 00:52:14,410
Edward Mehr: Yes, So, I think,
yeah, 3D printing was one of

928
00:52:14,410 --> 00:52:18,051
those examples where, overall as
an industry where the benefits

929
00:52:18,051 --> 00:52:21,635
were not as great as we thought.
There were some applications,

930
00:52:21,635 --> 00:52:25,045
though. If you look at within 3D
printing, you look at like

931
00:52:25,045 --> 00:52:27,993
dental. They found an
opportunity in an application

932
00:52:27,993 --> 00:52:30,940
that could provide 10 times
better for dentures and

933
00:52:30,940 --> 00:52:34,061
aligners. They're like 3D
printing can customize these

934
00:52:34,061 --> 00:52:37,125
things that were 10 times better
than the traditional

935
00:52:37,125 --> 00:52:40,419
alternative, where you have to
put like a little thing in

936
00:52:40,419 --> 00:52:44,118
someone's mouth and, and now you
can scan it and you're going to

937
00:52:44,118 --> 00:52:47,702
get it, right? so I think that
was the area that, and then you

938
00:52:47,702 --> 00:52:51,459
have to, mold, is expensive, you
have to manufacture it, now with

939
00:52:51,459 --> 00:52:55,158
3D printing, we can get aligners
or dentures really, easily. and

940
00:52:55,158 --> 00:52:58,510
it's also, looks much more
realistic with 3D printing, you

941
00:52:58,510 --> 00:53:01,920
can do all kinds of colors, and
you can make it really look

942
00:53:01,920 --> 00:53:05,215
realistic. even within 3D
printing, you look at the areas

943
00:53:05,215 --> 00:53:08,798
where there was opportunity for
improvement was 10 times, then

944
00:53:08,798 --> 00:53:12,382
they were successful. But any
area that you were just marginal

945
00:53:12,382 --> 00:53:14,000
improvement, it's dying out.

946
00:53:14,000 --> 00:53:17,756
Audrow Nash: How would you so
how did you find your area? You

947
00:53:17,756 --> 00:53:21,266
mentioned you wanted a big
market and then you wanted you

948
00:53:21,266 --> 00:53:25,084
looked for where there was a lot
of pain already in a sense is

949
00:53:25,084 --> 00:53:28,778
that kind of and then so from
that you find your problem and

950
00:53:28,778 --> 00:53:32,658
you look in the solution space
for what can be possibly solving

951
00:53:32,658 --> 00:53:36,537
that, and then you arrive at one
that might be feasible as a 10

952
00:53:36,537 --> 00:53:39,000
time solution or how do you
think of it?

953
00:53:39,000 --> 00:53:42,631
Edward Mehr: For us it was a
little bit, not as linear as you

954
00:53:42,631 --> 00:53:46,382
think, because I think there's
three components. There needs to

955
00:53:46,382 --> 00:53:50,073
be a big market, there needs to
be a solution, that also there

956
00:53:50,073 --> 00:53:53,585
needs to be enablers that you
can provide to that solution,

957
00:53:53,585 --> 00:53:57,157
that provides that ten times
increase in improvement, right?

958
00:53:57,157 --> 00:54:00,551
so for me, I was in additive
space. We saw the challenges

959
00:54:00,551 --> 00:54:04,301
we've seen with additive. So for
a couple of years, actually, I

960
00:54:04,301 --> 00:54:07,992
started thinking about, okay, if
you want autonomous and agile

961
00:54:07,992 --> 00:54:11,326
manufacturing that is not
product specific, and it's not

962
00:54:11,326 --> 00:54:15,136
part specific, it's not material
specific, where would that area

963
00:54:15,136 --> 00:54:18,827
be? I looked at post processing
and manufacturing. I looked at

964
00:54:18,827 --> 00:54:22,637
CNC machining. I looked at a lot
of different processes and then

965
00:54:22,637 --> 00:54:25,376
sheet metal is the one we
landed. Due to three

966
00:54:25,376 --> 00:54:28,948
combinations. One is, it was a
huge market size. There was a

967
00:54:28,948 --> 00:54:32,162
good amount of legacy work that
was being done. So the

968
00:54:32,162 --> 00:54:35,794
technology was almost ready. And
then my background and my co

969
00:54:35,794 --> 00:54:39,128
founder's background could
elevate it to the next stage,

970
00:54:39,128 --> 00:54:42,640
which was use robotics and
artificial intelligence to bring

971
00:54:42,640 --> 00:54:45,736
the cost of making this
manufacturing cell down, but

972
00:54:45,736 --> 00:54:49,486
also make it accurate, which was
the only shortcoming the sheet

973
00:54:49,486 --> 00:54:52,939
forming had. You couldn't make
accurate parts with it with

974
00:54:52,939 --> 00:54:56,451
incremental forming that was
done in academia. so for us it

975
00:54:56,451 --> 00:54:59,488
was a little bit of an
exploratory process. I don't

976
00:54:59,488 --> 00:55:03,000
think there's a, I don't think
there's a recipe that I can,

977
00:55:03,000 --> 00:55:04,000
Audrow Nash: Yeah. It's not
simple.

978
00:55:03,000 --> 00:55:06,412
Edward Mehr: Advertise. You just
got to iterate a bunch of

979
00:55:06,412 --> 00:55:09,529
processes and see where your
skill set can provide an

980
00:55:09,529 --> 00:55:13,118
enabler, but there's still a
huge market. and the jury's out.

981
00:55:13,118 --> 00:55:16,941
We'll see. I think we think it's
a big market we are going after,

982
00:55:16,941 --> 00:55:19,000
but we'll see. We'll see how we
do.

983
00:55:19,000 --> 00:55:23,264
Audrow Nash: you've made it this
far, which not to say that's

984
00:55:23,264 --> 00:55:27,599
success is not certain by any
means, but it's a good indicator

985
00:55:27,599 --> 00:55:32,003
I think, and especially 'cause
your valuation has been good and

986
00:55:32,003 --> 00:55:36,337
you have, I feel like a lot of
wind at your back already for a

987
00:55:36,337 --> 00:55:40,742
lot of manufacturing coming to
the US. So yeah, jury's out, but

988
00:55:40,742 --> 00:55:42,000
it's looking good.

989
00:55:42,000 --> 00:55:45,500
Edward Mehr: Looking good so
far. Yeah, I think, it's looking

990
00:55:45,500 --> 00:55:48,770
good so far, but it is tough. I
think that's probably the

991
00:55:48,770 --> 00:55:52,328
toughest portion. I'm actually,
I have an all hands coming up,

992
00:55:52,328 --> 00:55:55,770
which we're going to talk a
little bit about Makida's master

993
00:55:55,770 --> 00:55:59,328
plan. I'm planning to actually
put that on Twitter as well, or

994
00:55:59,328 --> 00:56:02,885
on X, so people can see. But I
think, Yeah, you need to really

995
00:56:02,885 --> 00:56:06,270
try to be, in early days, you
need to be ambitious, but not

996
00:56:06,270 --> 00:56:09,770
overly ambitious in terms of
tech, that the technology is not

997
00:56:09,770 --> 00:56:13,328
feasible. but you need to be as,
as ambitious as you can think

998
00:56:13,328 --> 00:56:16,541
the technology can take you. So
you have enough room for

999
00:56:16,541 --> 00:56:17,000
failure,

1000
00:56:17,000 --> 00:56:21,950
Audrow Nash: Yeah. A way, that
I've thought about it is, I've

1001
00:56:21,950 --> 00:56:26,982
seen many companies that have
ambitious plans, but their plans

1002
00:56:26,982 --> 00:56:31,933
rely on four, five, whatever
technologies that are all at the

1003
00:56:31,933 --> 00:56:36,964
cutting edge of being ready at
the same time. And that kind of

1004
00:56:36,964 --> 00:56:41,834
thing doesn't lead to terribly
good things. There's too much

1005
00:56:41,834 --> 00:56:46,541
risk in it, in my belief.
whereas the companies that I see

1006
00:56:46,541 --> 00:56:51,329
that are a lot more successful,
they pick almost always old

1007
00:56:51,329 --> 00:56:56,523
technology. Except for one area
where they're going to innovate,

1008
00:56:56,523 --> 00:57:01,635
and that's the one where they're
going to really push hard. And

1009
00:57:01,635 --> 00:57:06,343
they usually have some sort of
like great team that is all

1010
00:57:06,343 --> 00:57:10,887
leaders in this area. and
they're going to push that one

1011
00:57:10,887 --> 00:57:15,270
technology to the cusp of
whatever it is that they are

1012
00:57:15,270 --> 00:57:16,000
trying to

1013
00:57:16,000 --> 00:57:19,403
Edward Mehr: yep. Yep. It's two
sides of the thing, like you

1014
00:57:19,403 --> 00:57:22,806
need to have a big opportunity,
which provides you a lot of,

1015
00:57:22,806 --> 00:57:26,379
room to failure. But then on the
risk side, if you have five or

1016
00:57:26,379 --> 00:57:29,669
six or seven different risks in
technology, then also then

1017
00:57:29,669 --> 00:57:32,901
you're going to exhaust your
ability to fail, like you're

1018
00:57:32,901 --> 00:57:35,000
going to exhaust your failure,
quota,

1019
00:57:35,000 --> 00:57:36,000
Audrow Nash: Oh yeah.

1020
00:57:36,000 --> 00:57:39,244
Edward Mehr: so you need to just
make sure, yes, like physically

1021
00:57:39,244 --> 00:57:42,032
you're, you, you can get there
pretty fast. And this is

1022
00:57:42,032 --> 00:57:44,871
actually feasible technology
wise for sure. that's a big

1023
00:57:44,871 --> 00:57:48,014
portion of it. Like I said, in
our case it was like, robotics,

1024
00:57:48,014 --> 00:57:51,055
can they make the price of the
technology the system cheaper

1025
00:57:51,055 --> 00:57:54,147
using robotics? And that's why
we use off the shelf robotics,

1026
00:57:54,147 --> 00:57:57,290
as opposed to building our own
gantry system or our own custom

1027
00:57:57,290 --> 00:57:58,000
system, right?

1028
00:57:58,000 --> 00:58:01,469
Audrow Nash: And also it's less
to work on. It's one of those,

1029
00:58:01,469 --> 00:58:04,601
you're using old established
technology, not, you're not

1030
00:58:04,601 --> 00:58:06,000
building an arm yourself.

1031
00:58:06,000 --> 00:58:09,090
Edward Mehr: I don't need to
build a tool changer for it.

1032
00:58:09,090 --> 00:58:12,560
There are already tool changers
existing for it. I don't need to

1033
00:58:12,560 --> 00:58:15,380
build a calibration system for
it. There's already a

1034
00:58:15,380 --> 00:58:18,687
calibration system out there.
All I need to then focus on, if

1035
00:58:18,687 --> 00:58:21,452
I use robotics off the shelf, I
need to focus on my

1036
00:58:21,452 --> 00:58:24,867
intelligence. which is enabled
by these model buildings, right?

1037
00:58:24,867 --> 00:58:28,337
So now I have one big technical
risk, but it's a good risk. It's

1038
00:58:28,337 --> 00:58:31,536
a risk that we're seeing with
all the developments in AI is

1039
00:58:31,536 --> 00:58:34,843
constantly being improved and de
risked. So you're absolutely

1040
00:58:34,843 --> 00:58:38,205
right. You don't want to, you
don't want to build a company on

1041
00:58:38,205 --> 00:58:41,566
10 different risks, and hope,
hope that all of them will align

1042
00:58:41,566 --> 00:58:42,000
somehow.

1043
00:58:42,000 --> 00:58:43,000
Audrow Nash: definitely.

1044
00:58:43,000 --> 00:58:48,854
Audrow Nash: Now, not not saying
accepting venture capital money

1045
00:58:48,854 --> 00:58:54,159
is bad or anything, but I would
just be curious about your

1046
00:58:54,159 --> 00:58:59,646
thoughts on this. Do you think
you could do what you've done

1047
00:58:59,646 --> 00:59:05,317
without venture capital dollars
for this kind of thing or what

1048
00:59:05,317 --> 00:59:10,622
would the path be or how much
slower would it be? what are

1049
00:59:10,622 --> 00:59:13,000
your thoughts around this?

1050
00:59:13,000 --> 00:59:16,880
Edward Mehr: Venture also ties
into this conversation we just

1051
00:59:16,880 --> 00:59:20,505
had really well. Because what
does Venture do? Venture is

1052
00:59:20,505 --> 00:59:24,258
going after opportunities, the
colloquially say, each check

1053
00:59:24,258 --> 00:59:28,138
should be able to pay the fund
back. So they're taking a huge

1054
00:59:28,138 --> 00:59:31,700
risk, right? Especially in
earlier stages. They're like,

1055
00:59:31,700 --> 00:59:35,580
okay, it's 1 million check I'm
writing, needs to pay this 100

1056
00:59:35,580 --> 00:59:39,141
million fund back. That means
there's opportunity that 1

1057
00:59:39,141 --> 00:59:42,576
million can turn into 100
million for every investment

1058
00:59:42,576 --> 00:59:46,329
they make, because they probably
have one percent chance of

1059
00:59:46,329 --> 00:59:49,636
success. so by default, and
some, founders sometimes

1060
00:59:49,636 --> 00:59:53,452
complain about this and say, Oh,
why is venture investing in

1061
00:59:53,452 --> 00:59:57,014
these crazy ideas? Because
that's how their model works.

1062
00:59:57,014 --> 01:00:00,958
Their model works, but relies on
one success out of a hundred.

1063
01:00:00,958 --> 01:00:04,837
but when you're relying on one
success out of a hundred, that

1064
01:00:04,837 --> 01:00:07,000
needs to be very big upward,
very,

1065
01:00:07,000 --> 01:00:08,000
Audrow Nash: Lots of potential.

1066
01:00:08,000 --> 01:00:11,746
Edward Mehr: Lots of potential
and very opportunistic and very

1067
01:00:11,746 --> 01:00:15,552
huge ambitious, concept. So ties
into that conversation that we

1068
01:00:15,552 --> 01:00:18,814
had as well, like if you're
building a robotic company

1069
01:00:18,814 --> 01:00:22,500
that's incrementally improving
something, Venture wouldn't be

1070
01:00:22,500 --> 01:00:25,702
interested in it because it'd be
like, okay, how much

1071
01:00:25,702 --> 01:00:29,387
improvement, how much room is
there? so, ties very closely. I

1072
01:00:29,387 --> 01:00:32,649
think if you do have that
concept and you're convinced

1073
01:00:32,649 --> 01:00:36,516
yourself, very honestly that the
opportunity big and I can build

1074
01:00:36,516 --> 01:00:40,201
it The venture is good dollars.
I think venture comes in they

1075
01:00:40,201 --> 01:00:43,585
have expertise they have
network. So it's not just about

1076
01:00:43,585 --> 01:00:47,391
money It's about being able to
have a network of people who can

1077
01:00:47,391 --> 01:00:51,318
connect you to the right people.
For example, one of my investors

1078
01:00:51,318 --> 01:00:55,124
You know connected me to a lot
of folks at Stanford. I took a a

1079
01:00:55,124 --> 01:00:58,991
program at Stanford. It's an MBA
like program at Stanford. Helps

1080
01:00:58,991 --> 01:01:02,918
a lot with building the company,
hiring the right people. so I do

1081
01:01:02,918 --> 01:01:06,301
think venture is very helpful,
but not all companies are

1082
01:01:06,301 --> 01:01:09,744
venture fundable. if your
company's not venture fundable,

1083
01:01:09,744 --> 01:01:13,490
and you're going after venture
dollars, you're in for a lot of

1084
01:01:13,490 --> 01:01:17,055
pain. Because the expectations
are going to be high. You're

1085
01:01:17,055 --> 01:01:20,257
going to have really bad
interactions with your board

1086
01:01:20,257 --> 01:01:23,640
members. so that's something
that to be avoided. So make

1087
01:01:23,640 --> 01:01:27,386
sure, a lot of times it might
make sense to bootstrap your own

1088
01:01:27,386 --> 01:01:30,769
company. A lot of incremental
improvements make sense to

1089
01:01:30,769 --> 01:01:34,575
bootstrap it. try to see if you
can get some private funders or

1090
01:01:34,575 --> 01:01:38,442
your own money. Slowly build it
out. but if the concept requires

1091
01:01:38,442 --> 01:01:42,127
significant amount of capital
and there's a huge opportunity,

1092
01:01:42,127 --> 01:01:44,000
venture dollars can help a lot.

1093
01:01:44,000 --> 01:01:48,260
Audrow Nash: Gotcha. So it's
basically the upside. The

1094
01:01:48,260 --> 01:01:53,388
potential upside is a big way to
make this decision of if venture

1095
01:01:53,388 --> 01:01:58,200
capital might be appropriate.
And also, I suppose there's the

1096
01:01:58,200 --> 01:02:03,092
kind of implicit thing of you
need to have a good use for that

1097
01:02:03,092 --> 01:02:07,510
money. but if it's Incremental,
as you say, then there's

1098
01:02:07,510 --> 01:02:12,085
probably not that good of a use
for it. You can just do it

1099
01:02:12,085 --> 01:02:16,819
yourself. And if it is going to
be a hundred X or something,

1100
01:02:16,819 --> 01:02:21,000
then maybe venture could be a
very good fit for this.

1101
01:02:21,000 --> 01:02:24,237
Edward Mehr: Yeah, and to your
point, I think one point that

1102
01:02:24,237 --> 01:02:27,151
you brought up is pretty
important to do is that raise

1103
01:02:27,151 --> 01:02:30,227
the amount of dollars you need,
because it bites you back

1104
01:02:30,227 --> 01:02:33,464
afterwards as well, like in
terms of down rounds, right? So,

1105
01:02:33,464 --> 01:02:36,647
try to figure out how much
dollar you need and try to raise

1106
01:02:36,647 --> 01:02:39,507
as much as you need. Not,
necessarily more than that.

1107
01:02:39,507 --> 01:02:42,529
Because that's also another
challenging thing that comes

1108
01:02:42,529 --> 01:02:45,874
with the ventures. The higher
amount you raise, the higher the

1109
01:02:45,874 --> 01:02:49,058
expectation is going to be. So
make sure you're pacing your

1110
01:02:49,058 --> 01:02:51,000
business and progress based on
that.

1111
01:02:51,000 --> 01:02:53,000
Audrow Nash: Yeah. Makes sense.

1112
01:02:53,000 --> 01:02:57,199
Audrow Nash: So you guys. You're
70 people and you've been

1113
01:02:57,199 --> 01:03:01,688
growing really fast and you're
expecting to grow a good amount

1114
01:03:01,688 --> 01:03:05,887
more pretty quickly. how has
that experience been and what

1115
01:03:05,887 --> 01:03:09,000
have some of your lessons or
findings been?

1116
01:03:09,000 --> 01:03:12,834
Edward Mehr: Yeah, I think we're
chatting this with some of the,

1117
01:03:12,834 --> 01:03:16,668
other team members here. I think
the toughest part of building a

1118
01:03:16,668 --> 01:03:20,263
startup is actually still to
this day, it is not so much the

1119
01:03:20,263 --> 01:03:22,000
technology, it's the, people.

1120
01:03:22,000 --> 01:03:26,000
Audrow Nash: Definitely.

1121
01:03:26,000 --> 01:03:29,991
Edward Mehr: People are what the
company is. It's not the

1122
01:03:29,991 --> 01:03:33,911
technology. Tomorrow, the
technology can still be there,

1123
01:03:33,911 --> 01:03:38,182
but if all those people go, we
don't have much to build from.

1124
01:03:38,182 --> 01:03:42,592
so I think the biggest lesson is
hire well. and what that means

1125
01:03:42,592 --> 01:03:46,863
is, it needs to be people who
are passionate, mission driven,

1126
01:03:46,863 --> 01:03:51,064
and excited for that technology.
You don't want to just hire

1127
01:03:51,064 --> 01:03:54,914
smart people. If you want to
hire people that I have to

1128
01:03:54,914 --> 01:03:59,255
drive, excited about the mission
that you're after, the larger

1129
01:03:59,255 --> 01:04:03,036
problem you're trying to solve,
they can make light of

1130
01:04:03,036 --> 01:04:07,236
themselves, they can, in, in,
face of challenges and adverse

1131
01:04:07,236 --> 01:04:11,157
situations, they can still move
on. so there's a lot of,

1132
01:04:11,157 --> 01:04:15,357
basically most of my advice is
around. Choose very carefully

1133
01:04:15,357 --> 01:04:19,348
who you bring to your team,
especially as you're scaling,

1134
01:04:19,348 --> 01:04:23,409
because your company is the
people you hire. So that's one

1135
01:04:23,409 --> 01:04:27,889
condensed version of advice, but
there's a lot of branches there

1136
01:04:27,889 --> 01:04:32,160
to dive into. What does it mean
to be a good employee, a good

1137
01:04:32,160 --> 01:04:33,000
team member?

1138
01:04:33,000 --> 01:04:36,000
Audrow Nash: Do you want to
speak to it just for a bit?

1139
01:04:36,000 --> 01:04:39,816
Edward Mehr: Yeah, I think, If I
want to condense it, there's

1140
01:04:39,816 --> 01:04:43,381
probably or two or three
characteristics, that I've found

1141
01:04:43,381 --> 01:04:47,072
is very important. Number one is
mission driven. if you are

1142
01:04:47,072 --> 01:04:50,762
building a company that provides
10 x improvement, you want

1143
01:04:50,762 --> 01:04:54,265
people who are excited about
that 10 x improvement, that

1144
01:04:54,265 --> 01:04:57,830
vision that the, that, the
company built. You don't wanna

1145
01:04:57,830 --> 01:05:01,771
just bring experts in robotics
or machine learning that have no

1146
01:05:01,771 --> 01:05:05,399
connection to that. they don't
care if manufacturing world

1147
01:05:05,399 --> 01:05:09,152
becomes better or worse, right?
they're here to just provide

1148
01:05:09,152 --> 01:05:12,905
their expertise. so mission
drivenness is a big component of

1149
01:05:12,905 --> 01:05:16,596
it. I think probably the most
important portion. the second

1150
01:05:16,596 --> 01:05:20,537
one is obviously you want people
who are smart and that doesn't

1151
01:05:20,537 --> 01:05:24,165
mean necessarily they they know
and they're expert in that

1152
01:05:24,165 --> 01:05:28,105
field. It's good to be expert in
that field, but more important

1153
01:05:28,105 --> 01:05:31,671
than being an expert is that
somebody who can learn their

1154
01:05:31,671 --> 01:05:34,986
way. One thing I learned
throughout my career is that

1155
01:05:34,986 --> 01:05:38,614
every few years you have to
learn something new. there's a

1156
01:05:38,614 --> 01:05:42,617
new robotic framework, there's a
new modeling technique, there's

1157
01:05:42,617 --> 01:05:46,495
a new So yeah, and you need to
spend your weekend or whatever,

1158
01:05:46,495 --> 01:05:50,186
take a course on Coursera and
learn it and see what's going

1159
01:05:50,186 --> 01:05:52,000
on. To this date, I still do.

1160
01:05:52,000 --> 01:05:54,000
Audrow Nash: Oh, constantly.
Yeah, I'm sure.

1161
01:05:54,000 --> 01:05:57,673
Edward Mehr: Constantly, I was
like looking at the new, some

1162
01:05:57,673 --> 01:06:01,224
new changes in transformer
architecture over this weekend.

1163
01:06:01,224 --> 01:06:05,020
I'm just, I'm like, okay, what's
going on? I had to look at it

1164
01:06:05,020 --> 01:06:08,327
and learn it. so, that's
constant. So like smart, that

1165
01:06:08,327 --> 01:06:11,878
means that they learn fast. and
that's more important than

1166
01:06:11,878 --> 01:06:15,796
experts. and the last piece is,
I think, It's, you want doers. a

1167
01:06:15,796 --> 01:06:19,408
lot of experts and a lot of
smart people would love to chat

1168
01:06:19,408 --> 01:06:23,204
with you about how 10 million
ways a thing cannot be done. And

1169
01:06:23,204 --> 01:06:26,755
you're like, great, I don't want
to talk with you about 10

1170
01:06:26,755 --> 01:06:30,490
million ways that it cannot be
done. what is the way that can

1171
01:06:30,490 --> 01:06:34,102
be done? And let's try it. And
the problem is in our space,

1172
01:06:34,102 --> 01:06:37,286
there's a lot of unknown
unknowns. So you can't even

1173
01:06:37,286 --> 01:06:41,143
argue about it. You don't know,
like you might think, but until

1174
01:06:41,143 --> 01:06:44,878
you actually have done it, you
don't know. So you want doers,

1175
01:06:44,878 --> 01:06:48,367
you want people who have more
bias toward, once I have an

1176
01:06:48,367 --> 01:06:51,918
idea, I want to get it done as
soon as possible to see the

1177
01:06:51,918 --> 01:06:55,837
results as opposed to. Stick, be
stuck in analysis paralysis and

1178
01:06:55,837 --> 01:06:59,694
chat about it for days. I think
those three things, if somebody

1179
01:06:59,694 --> 01:07:03,122
has it, I think it's a core
things that, that means it's

1180
01:07:03,122 --> 01:07:06,000
probably a good person for early
stage startup.

1181
01:07:06,000 --> 01:07:10,146
Audrow Nash: That's funny. The
doer one is very funny. yeah,

1182
01:07:10,146 --> 01:07:14,293
it's really interesting. that
people can, I don't know, it's

1183
01:07:14,293 --> 01:07:17,955
like, it's like your
craftsmanship approach where you

1184
01:07:17,955 --> 01:07:21,894
have the metal and you're
gradually forming it but you're

1185
01:07:21,894 --> 01:07:26,179
looking and seeing what you do
to try to get it to the overall

1186
01:07:26,179 --> 01:07:30,463
shape that you want. It's like
you need to do that rather than

1187
01:07:30,463 --> 01:07:34,541
just perfect solution all at
once, because that's very hard

1188
01:07:34,541 --> 01:07:38,687
to do and come by and it might
just be wrong by the time you

1189
01:07:38,687 --> 01:07:40,000
implement it. Yeah.

1190
01:07:40,000 --> 01:07:42,797
Edward Mehr: I mean that
specific concept, when we were

1191
01:07:42,797 --> 01:07:45,898
first starting the company, I
was talking to a lot of robotic

1192
01:07:45,898 --> 01:07:48,746
experts. And we're like, Oh,
we're going to do this. And

1193
01:07:48,746 --> 01:07:51,593
everybody was like, Oh, the
robot will deflect. It's not

1194
01:07:51,593 --> 01:07:54,746
going to be accurate. It's not
going to form a part. And these

1195
01:07:54,746 --> 01:07:57,847
are experts in robotics, like
people built, segmented robots,

1196
01:07:57,847 --> 01:08:01,000
like the ones we're using. and
then it was funny. And then six

1197
01:08:01,000 --> 01:08:04,203
months later, we built the first
system and we formed the part.

1198
01:08:04,203 --> 01:08:07,356
And they're like, Yeah, I guess
it could be done. But like the

1199
01:08:07,356 --> 01:08:10,661
conversation before until it was
done, it was all about ways that

1200
01:08:10,661 --> 01:08:13,763
it cannot be done. and then once
it's done and it's shown and

1201
01:08:13,763 --> 01:08:16,915
it's done, then like people tear
in a leaf. but, but those are

1202
01:08:16,915 --> 01:08:19,712
the people you don't want in
your team. You almost want

1203
01:08:19,712 --> 01:08:22,814
people who are almost like naive
enough to try it. right? But

1204
01:08:22,814 --> 01:08:25,661
they have enough expert to back
and build a good system,

1205
01:08:25,661 --> 01:08:28,356
expertise to build a good
system, but they're not too

1206
01:08:28,356 --> 01:08:31,000
smart that prevents them to
actually take an action.

1207
01:08:31,000 --> 01:08:34,959
Audrow Nash: Yeah, it's funny
that we say too smart to do the

1208
01:08:34,959 --> 01:08:38,724
thing. I don't know. It's,
almost like if you're too smart

1209
01:08:38,724 --> 01:08:42,748
you see all the things but then
you can't do it. I don't think

1210
01:08:42,748 --> 01:08:46,577
it is that though. I would
imagine it's more fear of trying

1211
01:08:46,577 --> 01:08:50,471
and failing or something like
this. I think some people just

1212
01:08:50,471 --> 01:08:54,171
don't like uncertainty and
pushing into something and not

1213
01:08:54,171 --> 01:08:58,000
knowing if it will work or not
is terrifying in some sense.

1214
01:08:58,000 --> 01:09:01,712
Edward Mehr: Yeah. I mean I have
a lot of smart friends, I would

1215
01:09:01,712 --> 01:09:05,133
say very smart, is that because
they can play that chess in

1216
01:09:05,133 --> 01:09:08,439
their head. The reason I say
they're too smart is that if

1217
01:09:08,439 --> 01:09:12,034
they play chess with me, they
will always beat me because they

1218
01:09:12,034 --> 01:09:15,398
see 10 moves ahead, right? and I
almost say a person who's

1219
01:09:15,398 --> 01:09:18,994
moving, seeing 10 moves ahead
might not be a good fit for your

1220
01:09:18,994 --> 01:09:22,415
stage startups, right? Because,
the problem is in really in

1221
01:09:22,415 --> 01:09:25,373
chess, you might see 10 moves
ahead and it's a very

1222
01:09:25,373 --> 01:09:28,447
predictable environment. But
there's a lot of unknown

1223
01:09:28,447 --> 01:09:32,158
unknowns. so you want to almost,
it's almost as explore the map.

1224
01:09:32,158 --> 01:09:35,000
It's one of those game, strategy
games where you,

1225
01:09:35,000 --> 01:09:36,000
Audrow Nash: Fog of War,

1226
01:09:36,000 --> 01:09:38,995
Edward Mehr: open up, let's go
in that direction. And it's the

1227
01:09:38,995 --> 01:09:41,844
map opens up a little bit. And
then we can make a decision.

1228
01:09:41,844 --> 01:09:44,597
That's what really happens in
real world. It's not really

1229
01:09:44,597 --> 01:09:47,302
chess. So very, smart friends.
And we're like, this will

1230
01:09:47,302 --> 01:09:49,814
happen. Then this thing will
happen. Then that thing

1231
01:09:49,814 --> 01:09:52,664
happened. That's why it's not
possible. And we're like, but

1232
01:09:52,664 --> 01:09:55,562
the moment you do the first
state. There might be 10 options

1233
01:09:55,562 --> 01:09:58,460
open up that you didn't know you
couldn't think of today. in

1234
01:09:58,460 --> 01:10:01,503
chess, that's not the case, but
in real world, that's the case.

1235
01:10:01,503 --> 01:10:03,000
Because it's much more chaotic.

1236
01:10:03,000 --> 01:10:07,176
Audrow Nash: totally. Or if they
were playing chess, it's okay,

1237
01:10:07,176 --> 01:10:11,285
you have to pre decide your next
10 moves or whatever, and you

1238
01:10:11,285 --> 01:10:15,262
get to pick every single time
what your move will be looking

1239
01:10:15,262 --> 01:10:19,372
at the state of the board. You
would have a pretty significant

1240
01:10:19,372 --> 01:10:23,614
advantage . Anyways, I wanted to
get your thoughts on a bunch of

1241
01:10:23,614 --> 01:10:26,000
like contemporary technology
things.

1242
01:10:26,000 --> 01:10:29,750
Audrow Nash: Let's start with
humanoids. Tell me, about what

1243
01:10:29,750 --> 01:10:33,500
you think, what some of the
challenges will be, and then the

1244
01:10:33,500 --> 01:10:35,000
future that you imagine.

1245
01:10:35,000 --> 01:10:38,776
Edward Mehr: Yeah, I think
humanoid's an interesting

1246
01:10:38,776 --> 01:10:43,132
concept, I, like it. I think
humanoid's from the perspective

1247
01:10:43,132 --> 01:10:47,706
of generating excitement about
robotic space is as, close as it

1248
01:10:47,706 --> 01:10:51,700
gets in terms of oh, they look
like humans. They can do

1249
01:10:51,700 --> 01:10:56,056
everything we do. yeah. And
there are environments where you

1250
01:10:56,056 --> 01:11:00,703
certainly need humanoids, inside
homes, elderly care, might be a

1251
01:11:00,703 --> 01:11:04,987
good, area. Even in that area,
there, maybe there are other

1252
01:11:04,987 --> 01:11:09,343
form factors that are better,
but, there are certainly areas

1253
01:11:09,343 --> 01:11:13,917
where humanoids are very good.
And like I said, PR wise, it's a

1254
01:11:13,917 --> 01:11:17,838
very good story, because it
looks like a human walking

1255
01:11:17,838 --> 01:11:21,832
around, reasoning, talking, so
it's amazing in terms of

1256
01:11:21,832 --> 01:11:25,898
excitement. There are some
challenges in terms of people

1257
01:11:25,898 --> 01:11:30,254
are working on what does the
actuator actually need to be in

1258
01:11:30,254 --> 01:11:34,465
a humanoid? I have some of my
friends working on that. and

1259
01:11:34,465 --> 01:11:38,531
exciting developments there.
But, I think, We realized a

1260
01:11:38,531 --> 01:11:42,452
while back that human form is
not necessarily the most

1261
01:11:42,452 --> 01:11:47,026
optimized form for all tasks. in
manufacturing, we already know

1262
01:11:47,026 --> 01:11:51,383
robotic arms are much better.
They can be much more precise.

1263
01:11:51,383 --> 01:11:55,449
They can have much more
stiffness. They can apply higher

1264
01:11:55,449 --> 01:11:59,224
loads. So the amount of
flexibility you get over the

1265
01:11:59,224 --> 01:12:03,726
performance times performance is
actually pretty big. And it's

1266
01:12:03,726 --> 01:12:08,010
much better than a human for
form factor. So I would say in

1267
01:12:08,010 --> 01:12:12,366
areas like manufacturing. or
Logistics. A simpler form might

1268
01:12:12,366 --> 01:12:16,723
be more effective, and that
mostly, I think, is robotic art,

1269
01:12:16,723 --> 01:12:21,152
because you provide the same
kinematic freedom as a human, in

1270
01:12:21,152 --> 01:12:25,000
a much simpler, easier to
manufacture robotic system.

1271
01:12:25,000 --> 01:12:28,000
Audrow Nash: Oh, yeah,

1272
01:12:28,000 --> 01:12:31,849
Edward Mehr: so I think they're
going to be, eventually,

1273
01:12:31,849 --> 01:12:36,317
hopefully, we're going to have a
multitude of robot form factors.

1274
01:12:36,317 --> 01:12:40,579
and it's going to be different
for each application. Humanoids

1275
01:12:40,579 --> 01:12:44,909
will have a space, but I still
not, I'm still not convinced how

1276
01:12:44,909 --> 01:12:49,239
big that space is. as, as much
as people talk about it, I think

1277
01:12:49,239 --> 01:12:53,089
it's still pretty limited
because we long, too long ago,

1278
01:12:53,089 --> 01:12:57,419
we figured out that human form
is not necessarily the best form

1279
01:12:57,419 --> 01:12:59,000
for, many applications.

1280
01:12:59,000 --> 01:13:03,476
Audrow Nash: It's interesting.
The thing that stuck out to me

1281
01:13:03,476 --> 01:13:08,024
about your explanation is that
robot arms might be the optimal

1282
01:13:08,024 --> 01:13:12,353
form for a lot of things for
what we're doing. And that's a

1283
01:13:12,353 --> 01:13:16,829
really interesting idea. And it
makes good sense to me, being

1284
01:13:16,829 --> 01:13:21,158
that it's really just the most
simple way you can have some

1285
01:13:21,158 --> 01:13:25,340
sort of end effector moving
around doing something. So if

1286
01:13:25,340 --> 01:13:26,000
you need.

1287
01:13:26,000 --> 01:13:29,000
Edward Mehr: degrees of freedom
on the end effector, yeah. Yep.

1288
01:13:29,000 --> 01:13:33,308
Audrow Nash: So it's the
simplest way for that, and also

1289
01:13:33,308 --> 01:13:37,769
if you look of robotics at the
moment is that a lot of the

1290
01:13:37,769 --> 01:13:42,154
things that are very valuable
that we're doing are simple

1291
01:13:42,154 --> 01:13:47,000
mobile robots, where you're just
having a robot drive something

1292
01:13:47,000 --> 01:13:51,846
around moving from here to there
because we have a lot of those

1293
01:13:51,846 --> 01:13:56,615
types of problems and the state
of the technology is very well

1294
01:13:56,615 --> 01:14:01,462
suited for that type of problem
at the moment, and what I would

1295
01:14:01,462 --> 01:14:06,231
bet, connecting with your idea,
is that the next like 10 years

1296
01:14:06,231 --> 01:14:11,000
or something might be where we
see a lot of robotic arms doing

1297
01:14:11,000 --> 01:14:15,615
useful things in more spaces.
but, and they'll probably be a

1298
01:14:15,615 --> 01:14:20,231
more optimal form factor than a
humanoid which has a smaller

1299
01:14:20,231 --> 01:14:24,846
arm, or smaller arms, not as
much reach, not as much torque,

1300
01:14:24,846 --> 01:14:29,538
or stiffness, or these kinds of
things. I wonder, what do you

1301
01:14:29,538 --> 01:14:31,000
think of that idea?

1302
01:14:31,000 --> 01:14:34,608
Edward Mehr: No, I think, you're
absolutely right. I think

1303
01:14:34,608 --> 01:14:38,155
there's analogies in other
industries as well. I think of

1304
01:14:38,155 --> 01:14:41,950
robotic arms, they first got
deployed in 40s and 30s, roughly

1305
01:14:41,950 --> 01:14:45,496
around that time. I don't know
exactly when was the first

1306
01:14:45,496 --> 01:14:49,354
robotic arm was made, but it was
somewhere in Germany was used

1307
01:14:49,354 --> 01:14:53,273
for, in 30s or something or 20s.
so it has been in the industry

1308
01:14:53,273 --> 01:14:57,131
for a long time. So there's a
lot of legacy has been perfected

1309
01:14:57,131 --> 01:15:00,490
for a long time. Now there's
multiple vendors that are

1310
01:15:00,490 --> 01:15:03,850
competing, so it's almost
commoditized. so it's a very

1311
01:15:03,850 --> 01:15:07,520
good interface. The parallel of
this is to, we see a lot of

1312
01:15:07,520 --> 01:15:10,942
growth in NVIDIA, growth in
NVIDIA today. What was that

1313
01:15:10,942 --> 01:15:12,000
company? It was a

1314
01:15:12,000 --> 01:15:14,000
Audrow Nash: Just an enabler.
Yeah.

1315
01:15:14,000 --> 01:15:16,000
Edward Mehr: Gaming chip
company, right?

1316
01:15:16,000 --> 01:15:17,000
Audrow Nash: Uh-huh..

1317
01:15:17,000 --> 01:15:21,174
Edward Mehr: There is something
to be said why NVIDIA became

1318
01:15:21,174 --> 01:15:25,627
cornerstone of, computation for
neural networks. It's because of

1319
01:15:25,627 --> 01:15:29,384
the legacy. It's a chip that
already existed maybe for

1320
01:15:29,384 --> 01:15:33,350
another application, but it's a
very good fit. and it was

1321
01:15:33,350 --> 01:15:37,664
perfected. Now, maybe people are
thinking about other types of

1322
01:15:37,664 --> 01:15:41,768
chips. But the foundation was
built on a legacy system that

1323
01:15:41,768 --> 01:15:45,943
had other applications. Because
all the bugs, all the issues

1324
01:15:45,943 --> 01:15:49,769
have been ironed out, and it's
easy to use it, and it's

1325
01:15:49,769 --> 01:15:54,083
commoditized, and the price is
low, and you can just use it. I

1326
01:15:54,083 --> 01:15:58,118
think the same analogy works for
robotic arms, where these

1327
01:15:58,118 --> 01:16:02,432
systems have been perfected. We
have a very good supply chain,

1328
01:16:02,432 --> 01:16:06,467
Siemens makes a shit ton of,
electric motors, and Nabtesco

1329
01:16:06,467 --> 01:16:10,642
makes drives, and they have
perfected this supply chain. why

1330
01:16:10,642 --> 01:16:14,886
do we want to disrupt it for
another form factor? There needs

1331
01:16:14,886 --> 01:16:18,852
to be very good reasons, and I
think that the established

1332
01:16:18,852 --> 01:16:23,096
supply chain usually wins out if
the improvements are not ten

1333
01:16:23,096 --> 01:16:24,000
times better.

1334
01:16:24,000 --> 01:16:28,540
Audrow Nash: because they get a
lot of efficiency from the

1335
01:16:28,540 --> 01:16:32,844
amount of time they've been in
the space and the amount

1336
01:16:32,844 --> 01:16:37,384
products that they've pushed
out. Yeah, that's a very cool

1337
01:16:37,384 --> 01:16:42,236
perspective. I'll have to think
more about that. I like that a

1338
01:16:42,236 --> 01:16:46,776
lot. let's see, do you, what, so
one thing that I think is

1339
01:16:46,776 --> 01:16:51,081
especially interesting with
humanoids is you have large

1340
01:16:51,081 --> 01:16:56,090
valuations, like Figure, I don't
know, raising a billion dollars

1341
01:16:56,090 --> 01:17:00,864
or whatever it is. and they are
their own company raising the

1342
01:17:00,864 --> 01:17:05,638
money. I wonder if it will be
like autonomous cars, where you

1343
01:17:05,638 --> 01:17:10,021
see Cruise gets a billion
dollars and then its very hard

1344
01:17:10,021 --> 01:17:14,952
to deliver on the product in the
long run. But the thing that's

1345
01:17:14,952 --> 01:17:19,961
especially interesting to me is
then Tesla with their Tesla bot,

1346
01:17:19,961 --> 01:17:24,814
or Optimus, and the thing that's
really cool with that is that

1347
01:17:24,814 --> 01:17:29,353
they are their own captive,
market in a sense. Like it's a

1348
01:17:29,353 --> 01:17:34,206
very farsighted initiative. They
don't have to appease venture

1349
01:17:34,206 --> 01:17:38,667
capitalists, who want to return
on investment within some

1350
01:17:38,667 --> 01:17:42,894
timeframe of years. so I feel
like it's an interesting

1351
01:17:42,894 --> 01:17:47,198
dynamic, that they have been
able to fund this research

1352
01:17:47,198 --> 01:17:51,816
internally and maybe in on a
longer time horizon it will be

1353
01:17:51,816 --> 01:17:56,434
exciting or maybe just the
research can be applied to other

1354
01:17:56,434 --> 01:18:01,130
areas like maybe the top half of
a humanoid is very good for

1355
01:18:01,130 --> 01:18:05,669
upholstery in car manufacturing
for example. What are your

1356
01:18:05,669 --> 01:18:07,000
thoughts on this?

1357
01:18:07,000 --> 01:18:11,737
Edward Mehr: Yeah, it's a good
point you brought up. I think

1358
01:18:11,737 --> 01:18:16,000
all Elon companies should be
seen in a specific light.

1359
01:18:16,000 --> 01:18:18,731
Audrow Nash: I'm excited to hear
what that light is. Yeah, go

1360
01:18:18,731 --> 01:18:19,000
ahead.

1361
01:18:19,000 --> 01:18:22,945
Edward Mehr: I think one portion
of it is like he does things

1362
01:18:22,945 --> 01:18:26,697
that are very exciting, right?
Now he's, in charge of this

1363
01:18:26,697 --> 01:18:30,836
platform X. And, if you look at
a lot of the companies, he's, he

1364
01:18:30,836 --> 01:18:34,781
goes after that PR angle, right?
because I think that's what,

1365
01:18:34,781 --> 01:18:38,597
That's what generates enough
excitement for him to create a

1366
01:18:38,597 --> 01:18:42,607
big problem space that he can
solve. But the angle that people

1367
01:18:42,607 --> 01:18:46,682
usually forget, I think, about
Elon is that he does everything,

1368
01:18:46,682 --> 01:18:49,915
in my opinion, and I may be
wrong, but based on my

1369
01:18:49,915 --> 01:18:54,119
experience working at SpaceX and
interacting with a lot of people

1370
01:18:54,119 --> 01:18:58,000
who work with Elon even today,
Elon's number one goal always

1371
01:18:58,000 --> 01:19:01,493
has been becoming multi
planetary, and I think you can

1372
01:19:01,493 --> 01:19:05,373
see most of his companies in
light of that. Obviously SpaceX

1373
01:19:05,373 --> 01:19:09,448
is at the core of it, with the
Starship and Falcon 9, finding a

1374
01:19:09,448 --> 01:19:13,070
way where we can get to Mars and
other planets. Electric

1375
01:19:13,070 --> 01:19:16,821
propulsion is also another
reason that, that's pretty much

1376
01:19:16,821 --> 01:19:20,701
the only way you can travel in
other places. You cannot use,

1377
01:19:20,701 --> 01:19:24,000
necessarily a biofuels and
things that we have been

1378
01:19:24,000 --> 01:19:26,000
Audrow Nash: You run out or

1379
01:19:26,000 --> 01:19:29,769
Edward Mehr: out. Yes, electric
is probably the most, available

1380
01:19:29,769 --> 01:19:33,419
way of doing propulsion in those
places. and then you look at

1381
01:19:33,419 --> 01:19:36,590
Boring Company. If we go to
Mars, there's going to be

1382
01:19:36,590 --> 01:19:40,000
radiation out there, we need to
go down in the ground. So

1383
01:19:40,000 --> 01:19:44,000
Audrow Nash: Aha, that's wild. I
never connected all this. Okay.

1384
01:19:44,000 --> 01:19:47,815
Edward Mehr: you need to go down
and do that. neural link is

1385
01:19:47,815 --> 01:19:51,693
also, we're going to come back
to it, but it's, about kind of

1386
01:19:51,693 --> 01:19:55,062
high bandwidth control
interface. and then I think of

1387
01:19:55,062 --> 01:19:59,068
humanoids also an easy interface
for astronauts. if you build a

1388
01:19:59,068 --> 01:20:03,136
system for humans, then you can
replace them with humanoids for,

1389
01:20:03,136 --> 01:20:07,015
outer space applications. So a
lot of times, a lot of bend to

1390
01:20:07,015 --> 01:20:10,829
Elon's companies are around, I
think he's building different

1391
01:20:10,829 --> 01:20:14,453
components of that larger term
goal that he has, which is

1392
01:20:14,453 --> 01:20:18,268
becoming multi planetary. And
he's being very creative about

1393
01:20:18,268 --> 01:20:22,082
it, like finding business use
cases on earth that can propel

1394
01:20:22,082 --> 01:20:25,706
this. But I do think he always
had that bend toward multi

1395
01:20:25,706 --> 01:20:29,457
planetary and that's why he
starts from there basically. He

1396
01:20:29,457 --> 01:20:32,000
looks at what do I need on Mars
and then

1397
01:20:32,000 --> 01:20:35,786
Audrow Nash: aha. big build
backwards and what do you need

1398
01:20:35,786 --> 01:20:39,312
to build here on Earth. So
that's that Figure. I don't

1399
01:20:39,312 --> 01:20:43,294
know. I don't know enough to
know, but I think Elon companies

1400
01:20:43,294 --> 01:20:46,950
at least are unique in that
perspective. Yeah, that is a

1401
01:20:46,950 --> 01:20:50,801
very interesting way to frame
it. And it makes sense. I had

1402
01:20:50,801 --> 01:20:54,000
never connected those dots. I
love it. Hell yeah.

1403
01:20:54,000 --> 01:20:57,529
Audrow Nash: And then so now
moving on to AI. when you guys,

1404
01:20:57,529 --> 01:21:01,176
when you, when we started, you
were saying AI, you're using AI

1405
01:21:01,176 --> 01:21:05,000
and I think AI has become a good
bit of a like buzzword marketing

1406
01:21:05,000 --> 01:21:08,471
thing. Like everyone wants to
say AI when they mean machine

1407
01:21:08,471 --> 01:21:09,000
learning,

1408
01:21:09,000 --> 01:21:10,000
Edward Mehr: Yeah.

1409
01:21:10,000 --> 01:21:14,014
Audrow Nash: which is
optimization from my

1410
01:21:14,014 --> 01:21:19,939
perspective, like regression or
things like this. And, so from

1411
01:21:19,939 --> 01:21:25,672
my perspective of what you guys
are doing, it's not chatGPT,

1412
01:21:25,672 --> 01:21:31,215
it's not, you're using neural
networks, but I, don't know,

1413
01:21:31,215 --> 01:21:37,044
tell me about your thoughts on
AI and then we'll go into more

1414
01:21:37,044 --> 01:21:38,000
specifics.

1415
01:21:38,000 --> 01:21:41,971
Edward Mehr: yeah. It's
interesting. the definition of

1416
01:21:41,971 --> 01:21:43,000
AI has changed

1417
01:21:43,000 --> 01:21:44,000
Audrow Nash: Oh yeah,

1418
01:21:44,000 --> 01:21:48,510
Edward Mehr: over time, right? I
think, like I said, if you look

1419
01:21:48,510 --> 01:21:52,174
at back in 60s and 70s, like
rule based systems were

1420
01:21:52,174 --> 01:21:55,979
considered artificial
intelligence. then when I was in

1421
01:21:55,979 --> 01:22:00,348
school in 2004, we were talking
about kind of tree search as a

1422
01:22:00,348 --> 01:22:04,012
AI system where like you
basically, like rolling out

1423
01:22:04,012 --> 01:22:08,311
future in multiple steps, you
optimize and then you get that.

1424
01:22:08,311 --> 01:22:12,187
And then with the data, the
definition moved toward ML.

1425
01:22:12,187 --> 01:22:16,133
Basically, can I find the
pattern? In, in a set of data,

1426
01:22:16,133 --> 01:22:19,938
the same way human finds a
pattern, like you look at a

1427
01:22:19,938 --> 01:22:24,025
picture and the pattern of
colors and lines tells you it's

1428
01:22:24,025 --> 01:22:28,464
a dog. and that became like the
state of the art for AI. Now, I

1429
01:22:28,464 --> 01:22:32,269
think with transformers, that
again, change into large

1430
01:22:32,269 --> 01:22:36,356
language models, became the
state of the art of AI because

1431
01:22:36,356 --> 01:22:40,866
you could talk to it and it will
talk to you like, like a human.

1432
01:22:40,866 --> 01:22:44,601
But the core of large language
models that are GPT or

1433
01:22:44,601 --> 01:22:49,040
transformers, is the same as, to
me, for me, it's the same as a

1434
01:22:49,040 --> 01:22:52,000
more simpler neural network,
but, and then

1435
01:22:52,000 --> 01:22:53,000
Audrow Nash: It just

1436
01:22:52,000 --> 01:22:55,122
Edward Mehr: is the same as, and
each node of those is basically

1437
01:22:55,122 --> 01:22:56,000
a regression mode.

1438
01:22:56,000 --> 01:22:58,000
Audrow Nash: Yeah,

1439
01:22:58,000 --> 01:23:02,829
Edward Mehr: so they're all
almost like maturity of the same

1440
01:23:02,829 --> 01:23:07,256
system, all the way even to
large language models. so I

1441
01:23:07,256 --> 01:23:11,763
categorize them as the same
technology, and AI being the

1442
01:23:11,763 --> 01:23:16,512
term that we always apply to the
most advanced cutting edge

1443
01:23:16,512 --> 01:23:21,261
portion of that. so yes, we do
use machine learning. we are

1444
01:23:21,261 --> 01:23:26,412
using graph, neural networks. So
closer, and then with this next

1445
01:23:26,412 --> 01:23:29,873
project, we are using
transformers, to do a

1446
01:23:29,873 --> 01:23:31,000
multimodal, of

1447
01:23:31,000 --> 01:23:33,000
Audrow Nash: So you can have a
context remembered

1448
01:23:33,000 --> 01:23:35,781
Edward Mehr: You can have a
context and you can also do,

1449
01:23:35,781 --> 01:23:38,911
we're talking about multimodal
models now where you can feed it

1450
01:23:38,911 --> 01:23:41,692
an image of a CAD of a part, but
you can also feed it an

1451
01:23:41,692 --> 01:23:44,871
instruction of what accuracy you
want. And then come up with the

1452
01:23:44,871 --> 01:23:47,503
robot actions. So it's a
multimodal model that we are

1453
01:23:47,503 --> 01:23:48,000
looking at

1454
01:23:48,000 --> 01:23:51,073
Audrow Nash: That's super cool.
Okay, I got to learn more about

1455
01:23:51,073 --> 01:23:52,000
that kind of thing.

1456
01:23:52,000 --> 01:23:55,913
Edward Mehr: So that is closer
to what is happening with

1457
01:23:55,913 --> 01:23:59,477
ChatGPT. Now, I think the
definition of the AI will

1458
01:23:59,477 --> 01:24:03,390
progress every time we beat the
boundary. Now, is it the

1459
01:24:03,390 --> 01:24:07,792
Terminator or the, Skynet? Not
yet. in pop culture, that's what

1460
01:24:07,792 --> 01:24:11,845
people think of AI. But, if you
go into academic text, the

1461
01:24:11,845 --> 01:24:15,898
definition has changed over
time. So I don't know, I don't

1462
01:24:15,898 --> 01:24:20,230
know what to call actually. like
you can talk about, go with a

1463
01:24:20,230 --> 01:24:24,492
pop culture reference. I don't
think anybody's there yet. but

1464
01:24:24,492 --> 01:24:28,266
if you go with the LLM, then you
can talk about, okay,

1465
01:24:28,266 --> 01:24:32,248
transformers are maybe AI,
building components AI. We are

1466
01:24:32,248 --> 01:24:36,301
modeling using transformers. So
it's a question of how you

1467
01:24:36,301 --> 01:24:37,000
define it.

1468
01:24:37,000 --> 01:24:41,688
Audrow Nash: yeah, for sure.
Yeah, it's been funny that it's

1469
01:24:41,688 --> 01:24:46,141
whatever is, I don't know, I
guess AGI Artificial General

1470
01:24:46,141 --> 01:24:50,672
Intelligence is like the new
thing everyone's trying to do

1471
01:24:50,672 --> 01:24:55,516
and that's the holy grail of all
this AI stuff but it's, funny

1472
01:24:55,516 --> 01:24:59,969
that it's oh, artificial
intelligence will never get this

1473
01:24:59,969 --> 01:25:04,813
and then it completes in chess
and then it. competes and beats

1474
01:25:04,813 --> 01:25:09,578
the, top guy or the top person
and go, and then it's oh, it's

1475
01:25:09,578 --> 01:25:14,578
not AI yet. Every, one of these
ones we redefine ai. yeah. I see

1476
01:25:14,578 --> 01:25:19,266
what you mean. Where we change
the definition. What are your

1477
01:25:19,266 --> 01:25:23,172
thoughts on ChatGPT, or large
language models more

1478
01:25:23,172 --> 01:25:27,000
specifically and where they're
useful, I suppose?

1479
01:25:27,000 --> 01:25:30,583
Edward Mehr: Yeah, so I think
The way I think about large

1480
01:25:30,583 --> 01:25:34,606
language models, and I'm not an
expert in the field, I have very

1481
01:25:34,606 --> 01:25:38,314
surface knowledge, in that
field. I think there's something

1482
01:25:38,314 --> 01:25:42,211
special about them. obviously
they're language models, so they

1483
01:25:42,211 --> 01:25:45,920
can just talk, right? But I
think they're capturing the way

1484
01:25:45,920 --> 01:25:49,691
we think, because if you can
say, I think of language as the

1485
01:25:49,691 --> 01:25:53,463
way we are encoding a human
knowledge and thought process, I

1486
01:25:53,463 --> 01:25:57,486
think mastering that is as close
as possible we have figured out

1487
01:25:57,486 --> 01:26:00,754
to this date into, into,
thinking. So then there are

1488
01:26:00,754 --> 01:26:04,463
areas in, manufacturing where
that directly applies, right?

1489
01:26:04,463 --> 01:26:08,360
and mostly is around interface
improvement. if I'm programming

1490
01:26:08,360 --> 01:26:12,006
a robot to do certain things,
can I just talk to it? I can

1491
01:26:12,006 --> 01:26:15,966
just get to a point where I say,
Okay, should I code up forming

1492
01:26:15,966 --> 01:26:19,800
this geometry? Or can I just
say, form this CAD and then say,

1493
01:26:19,800 --> 01:26:23,823
Okay. Scan it and they will tell
me what is the metric and I say

1494
01:26:23,823 --> 01:26:27,720
can you improve it by 10 percent
and then they will do another

1495
01:26:27,720 --> 01:26:31,429
round of forming. So the direct
area there is like removing

1496
01:26:31,429 --> 01:26:34,949
friction in an interface with
robotics. And I think that

1497
01:26:34,949 --> 01:26:38,846
definitely is on our horizon as
well. We're looking into that.

1498
01:26:38,846 --> 01:26:42,366
Maybe not first order problems
we want to solve but it's

1499
01:26:42,366 --> 01:26:44,000
definitely in our horizon.

1500
01:26:44,000 --> 01:26:46,000
Audrow Nash: Eventually. Yeah.

1501
01:26:46,000 --> 01:26:50,418
Edward Mehr: But with multimodal
models I think we are getting,

1502
01:26:50,418 --> 01:26:54,415
much closer to that being
applied in manufacturing. I was

1503
01:26:54,415 --> 01:26:58,132
following one of our great
friends, Pieter Abbeel, at

1504
01:26:58,132 --> 01:27:02,339
Covariant, they released RFM1
where it's a multimodal model.

1505
01:27:02,339 --> 01:27:06,757
You can just say, okay, here's a
picture of a bin and here's an

1506
01:27:06,757 --> 01:27:11,105
instruction, pick up a banana
from it. And the robot will, the

1507
01:27:11,105 --> 01:27:15,383
model outputs robot actions on
the other end. We'll pick up a

1508
01:27:15,383 --> 01:27:19,871
banana from that image. so now.
Now we're getting not just text,

1509
01:27:19,871 --> 01:27:24,079
we're using the same component,
which are transformers, same

1510
01:27:24,079 --> 01:27:27,795
building block, but creating
these multi modal, multi

1511
01:27:27,795 --> 01:27:31,933
interface, multi medium, models
that can input image, still

1512
01:27:31,933 --> 01:27:36,000
frame, large language, and then
outputting the same thing.

1513
01:27:36,000 --> 01:27:37,646
Audrow Nash: You just convert it
to data and it will do something

1514
01:27:37,646 --> 01:27:38,000
with the data.

1515
01:27:38,000 --> 01:27:41,503
Edward Mehr: And they're
combining image and natural

1516
01:27:41,503 --> 01:27:45,342
language in a way that, the
brain does. and that's pretty

1517
01:27:45,342 --> 01:27:49,181
exciting. I think that has a
huge potential to interrupt,

1518
01:27:49,181 --> 01:27:51,000
disrupt, the robotic space.

1519
01:27:51,000 --> 01:27:55,578
Audrow Nash: Gotcha. One of the
things that was interesting to

1520
01:27:55,578 --> 01:28:00,303
me about especially working with
ChatGPT and maybe this won't be

1521
01:28:00,303 --> 01:28:05,102
as much of an issue for robotics
companies because the task space

1522
01:28:05,102 --> 01:28:09,089
could be smaller but there's a
lot of weirdness around

1523
01:28:09,089 --> 01:28:13,667
predictable results where it'll
just hallucinate something and

1524
01:28:13,667 --> 01:28:17,949
they are getting better like
specifically in my experience

1525
01:28:17,949 --> 01:28:22,453
lately, Claude3 is way better,
at hallucinating far less than

1526
01:28:22,453 --> 01:28:27,031
ChatGPT. That has been really
cool. That actually might be the

1527
01:28:27,031 --> 01:28:31,313
first one that I can actually
use rather than just getting

1528
01:28:31,313 --> 01:28:35,227
really early ideas about
something or like very basic

1529
01:28:35,227 --> 01:28:39,657
coding help. that one might
actually be, I might actually be

1530
01:28:39,657 --> 01:28:44,013
able to enlist it for like
podcast related tasks. But, what

1531
01:28:44,013 --> 01:28:48,000
do you think about the
hallucinations of these models?

1532
01:28:48,000 --> 01:28:51,376
Edward Mehr: Yeah, this is a
new, not a new thing, right?

1533
01:28:51,376 --> 01:28:54,753
we're from a robotic system. We
always, we've worked with

1534
01:28:54,753 --> 01:28:58,544
deterministic systems where same
input gives you the same output

1535
01:28:58,544 --> 01:29:02,038
every time. But we have been
dealing with humans and humans

1536
01:29:02,038 --> 01:29:05,474
are very non deterministic,
right? And they are a big part

1537
01:29:05,474 --> 01:29:08,850
of manufacturing process, right?
Humans are a big part of

1538
01:29:08,850 --> 01:29:12,523
manufacturing and the way we got
around that is through checks

1539
01:29:12,523 --> 01:29:16,077
and balances. Right, and we say,
okay, if the output of this

1540
01:29:16,077 --> 01:29:19,749
result needs to verify this way,
as long as it's passing these

1541
01:29:19,749 --> 01:29:23,422
verifications is good enough.
Because in reality, we could not

1542
01:29:23,422 --> 01:29:27,094
rely on human output. One day
I'm sick and my, mistake, make a

1543
01:29:27,094 --> 01:29:30,767
mistake. And the next day I'm, I
am, I might outperform what I

1544
01:29:30,767 --> 01:29:34,380
usually do, right? So we already
have built systems that rely

1545
01:29:34,380 --> 01:29:37,282
with, that, deals with
ambiguity. So I'm not that

1546
01:29:37,282 --> 01:29:39,000
concerned about it, honestly.

1547
01:29:39,000 --> 01:29:42,506
Audrow Nash: Yeah you just have
your checks and verifications I

1548
01:29:42,506 --> 01:29:45,734
think that's great because
you're right, you don't need to

1549
01:29:45,734 --> 01:29:49,184
give it total control and have
it do crazy. if you were having

1550
01:29:49,184 --> 01:29:52,523
a robot choose actions from a
language model or maybe like a

1551
01:29:52,523 --> 01:29:55,640
language action model or
something like this, or a large

1552
01:29:55,640 --> 01:29:59,146
action model, you could just run
it through a path planner at a

1553
01:29:59,146 --> 01:30:02,374
very high level with a very low
fidelity simulation really

1554
01:30:02,374 --> 01:30:05,824
quickly to make sure it's not
going to collide with stuff. And

1555
01:30:05,824 --> 01:30:09,052
that would be a very simple
check on the proposed actions,

1556
01:30:09,052 --> 01:30:11,000
which is actually a very cool
idea.

1557
01:30:11,000 --> 01:30:14,123
Edward Mehr: No, you're
absolutely right. And then the

1558
01:30:14,123 --> 01:30:17,477
other piece is, I think, at
least the way we thought about

1559
01:30:17,477 --> 01:30:20,484
it, is that once you want
repeatability. You can use

1560
01:30:20,484 --> 01:30:23,606
neural networks and machine
learning and AI to get, to

1561
01:30:23,606 --> 01:30:27,250
develop something and say, okay,
this is the, like in our case,

1562
01:30:27,250 --> 01:30:31,008
for example, you can ask a model
to just say, what is the process

1563
01:30:31,008 --> 01:30:34,709
parameters? What path the robot
needs to take to form this part,

1564
01:30:34,709 --> 01:30:38,295
this aircraft wing, below one
millimeter accuracy. And it will

1565
01:30:38,295 --> 01:30:41,938
come up with that. And maybe you
make a mistake and you have to

1566
01:30:41,938 --> 01:30:45,408
fine tune the model. But then
the moment you want to make 10

1567
01:30:45,408 --> 01:30:48,877
of them, then close it. Don't
ask every time from chat, from

1568
01:30:48,877 --> 01:30:52,000
your model to how to make the
same thing over and over

1569
01:30:52,000 --> 01:30:53,000
Audrow Nash: You already got

1570
01:30:53,000 --> 01:30:56,345
Edward Mehr: close it. So you
can also create these, almost

1571
01:30:56,345 --> 01:30:59,747
silos and be like, okay, in
development, I'll use the models

1572
01:30:59,747 --> 01:31:03,093
then, but in production, I'll
lock it down. at least that's

1573
01:31:03,093 --> 01:31:06,438
how we were thinking about it.
For example, for our case is

1574
01:31:06,438 --> 01:31:09,613
that we don't want, once we
figured out how to form this

1575
01:31:09,613 --> 01:31:12,902
wing, we don't want to ask the
AI anymore. We just want to

1576
01:31:12,902 --> 01:31:15,000
replicate the recipe it came up
with.

1577
01:31:15,000 --> 01:31:18,631
Audrow Nash: Yeah. I like that a
lot. Any, thoughts on the form

1578
01:31:18,631 --> 01:31:21,744
that these checks and
verifications for LLMs, like I'm

1579
01:31:21,744 --> 01:31:25,144
just thinking of like unit and
integration testing for this

1580
01:31:25,144 --> 01:31:28,660
kind of thing, and maybe some
like high level safeguards, but

1581
01:31:28,660 --> 01:31:32,406
any thoughts on this? Cause it's
a really cool idea. This is what

1582
01:31:32,406 --> 01:31:35,000
I'm going to think a lot more
about, I think.

1583
01:31:35,000 --> 01:31:38,159
Edward Mehr: Yeah, I think you
nailed it in most respects. As

1584
01:31:38,159 --> 01:31:41,110
long as it's not going to
destroy the system, I think you

1585
01:31:41,110 --> 01:31:44,269
can let the robot do what it
needs to do. As long as it's not

1586
01:31:44,269 --> 01:31:47,168
destroying itself. it's that
means, maybe don't go above

1587
01:31:47,168 --> 01:31:50,482
certain currents, don't run into
yourself. You just talked about

1588
01:31:50,482 --> 01:31:51,000
collision.

1589
01:31:51,000 --> 01:31:56,096
Audrow Nash: So you give it a,
safe like parameter, a safe set

1590
01:31:56,096 --> 01:31:57,000
of actions,

1591
01:31:57,000 --> 01:32:00,024
Edward Mehr: yes, and that's a
much bigger space than what you

1592
01:32:00,024 --> 01:32:03,000
could think of, but at least the
robot doesn't damage itself.

1593
01:32:03,000 --> 01:32:04,000
Audrow Nash: Yeah.

1594
01:32:04,000 --> 01:32:08,366
Edward Mehr: And I think as long
as it's not damaging itself, I

1595
01:32:08,366 --> 01:32:11,000
think you can probably let it
explore.

1596
01:32:11,000 --> 01:32:12,000
Audrow Nash: huh.

1597
01:32:12,000 --> 01:32:14,000
Edward Mehr: At least that's how
I'm thinking about it.

1598
01:32:14,000 --> 01:32:17,360
Audrow Nash: Yeah, that makes a
lot of sense to me. What do you

1599
01:32:17,360 --> 01:32:20,240
think about the role of
simulation? In all of this and

1600
01:32:20,240 --> 01:32:22,000
sim2real and everything like
this

1601
01:32:22,000 --> 01:32:25,685
Edward Mehr: Yeah. That's
another interesting area, right?

1602
01:32:25,685 --> 01:32:29,752
like you can let these robots to
be in the simulation world. For

1603
01:32:29,752 --> 01:32:33,628
us, it's a little bit tougher
because the main reason we went

1604
01:32:33,628 --> 01:32:36,742
to AI based models is that
physics simulation was

1605
01:32:36,742 --> 01:32:40,427
complicated, and tough to do.
but that's another branch of

1606
01:32:40,427 --> 01:32:44,303
work we're doing at the moment.
we actually did this proposal

1607
01:32:44,303 --> 01:32:48,179
with, Lawrence Livermore at the
National Lab a while back and

1608
01:32:48,179 --> 01:32:52,246
trying to create a simulation of
our process that is on GPUs and

1609
01:32:52,246 --> 01:32:55,995
it's faster. so I think there's
opportunity there to create

1610
01:32:55,995 --> 01:33:00,125
augmented data, even if, as long
as the simulation is fast, maybe

1611
01:33:00,125 --> 01:33:03,874
it's not as accurate, but if
it's fast, then I think it can

1612
01:33:03,874 --> 01:33:07,496
be helpful, right? Because you
get like roughly the right

1613
01:33:07,496 --> 01:33:11,499
thing, and then you can do a lot
of checks and balances, and it

1614
01:33:11,499 --> 01:33:15,566
generates you a lot of synthetic
data to also improve the model.

1615
01:33:15,566 --> 01:33:19,060
As long as directionally, it's
correct. If the model is

1616
01:33:19,060 --> 01:33:23,000
directionally not correct, then
it's going to be really tough.

1617
01:33:23,000 --> 01:33:25,060
Audrow Nash: what do you mean
directionally if it's doing

1618
01:33:25,060 --> 01:33:26,000
completely the wrong thing

1619
01:33:26,000 --> 01:33:29,437
Edward Mehr: yeah, completely
wrong. It's maybe it doesn't

1620
01:33:29,437 --> 01:33:32,756
accurately figure out what's
going to happen in the real

1621
01:33:32,756 --> 01:33:36,548
world, but you know that if you
increase the speed, it will, the

1622
01:33:36,548 --> 01:33:40,163
outcome also moves in the right
direction. That is similar to

1623
01:33:40,163 --> 01:33:42,000
what happens in the real world.

1624
01:33:42,000 --> 01:33:46,409
Audrow Nash: Yeah, one work that
I saw quite a while ago I'm sure

1625
01:33:46,409 --> 01:33:50,818
you guys are aware of stuff like
this, but they were using a very

1626
01:33:50,818 --> 01:33:55,092
high fidelity simulator and they
were using machine learning to

1627
01:33:55,092 --> 01:33:59,162
try to link one frame to the
next of complex objects. And so

1628
01:33:59,162 --> 01:34:03,435
the simulator, they ran it super
slow in super high resolution,

1629
01:34:03,435 --> 01:34:07,302
very tiny timestamp, all these
things like this. But they

1630
01:34:07,302 --> 01:34:11,575
learned the transitions and then
within a space of. Things that

1631
01:34:11,575 --> 01:34:15,849
it had seen it was a very fast
simulator because you just had a

1632
01:34:15,849 --> 01:34:19,851
pass through the feed forward
network effectively. and that

1633
01:34:19,851 --> 01:34:23,921
kind of thing seems really cool.
And if you apply a check on

1634
01:34:23,921 --> 01:34:28,127
that, which says, Hey, are you
would like super low resolution

1635
01:34:28,127 --> 01:34:32,400
version of the simulation? Let
me make sure that things are not

1636
01:34:32,400 --> 01:34:36,131
just completely wonky,
everything's shooting across the

1637
01:34:36,131 --> 01:34:40,201
screen or whatever. Then you
could probably use that kind of

1638
01:34:40,201 --> 01:34:44,000
thing for learning, at a fast
rate, which is quite cool.

1639
01:34:44,000 --> 01:34:47,325
Edward Mehr: No, I agree. I
think that, that building that

1640
01:34:47,325 --> 01:34:50,765
kind of a proxy. you're almost
building an empirical machine

1641
01:34:50,765 --> 01:34:53,976
learning based proxy model for
the simulation using that

1642
01:34:53,976 --> 01:34:57,645
technique. which is, fantastic.
I think that's a good way of, if

1643
01:34:57,645 --> 01:35:00,856
you cannot make a lot of data in
the real world, use the

1644
01:35:00,856 --> 01:35:04,067
simulation to augment it. And as
long as, like you said,

1645
01:35:04,067 --> 01:35:07,621
directionally is correct. It's
not going to do, if I move this

1646
01:35:07,621 --> 01:35:11,233
to the right in the simulation,
it's going to move right in the

1647
01:35:11,233 --> 01:35:14,788
real world as well. As long as
that directionality is there, I

1648
01:35:14,788 --> 01:35:18,457
think you can generate a lot of
data and learn a lot of patterns

1649
01:35:18,457 --> 01:35:21,783
from it for your model, and then
retrain it or do transfer

1650
01:35:21,783 --> 01:35:24,993
learning in the real world to
then fine tune it based on

1651
01:35:24,993 --> 01:35:27,000
what's happening in the real
world.

1652
01:35:27,000 --> 01:35:32,000
Audrow Nash: Oh, yeah. changing
topic a little bit.

1653
01:35:32,000 --> 01:35:36,261
Audrow Nash: Looking out at the
manufacturing space? What are

1654
01:35:36,261 --> 01:35:40,592
some things that are top of mind
for you, predictions maybe in

1655
01:35:40,592 --> 01:35:44,783
the next five years or so? or
where are you watching? what's

1656
01:35:44,783 --> 01:35:49,044
interesting to you, looking out
at the future and considering

1657
01:35:49,044 --> 01:35:51,000
our world and manufacturing?

1658
01:35:51,000 --> 01:35:55,378
Edward Mehr: Yeah, like we said,
there's a lot of wind for, on

1659
01:35:55,378 --> 01:35:59,544
shoring manufacturing. I think
there's going to be a lot of

1660
01:35:59,544 --> 01:36:03,568
policy, both investment and the
government investment and

1661
01:36:03,568 --> 01:36:07,734
private investment in that
space. But what I think is, it's

1662
01:36:07,734 --> 01:36:11,617
interesting, there's two camps.
There's a camp of Let's

1663
01:36:11,617 --> 01:36:15,713
replicate what we did in the
1960s and 1940s and World War

1664
01:36:15,713 --> 01:36:19,455
II. Let's bring all that scale
back. oh, let's create

1665
01:36:19,455 --> 01:36:23,691
foundries, and let's create,
machine shops, and let's make a

1666
01:36:23,691 --> 01:36:27,787
lot of them, right? And there
is, I think, a camp of Let's

1667
01:36:27,787 --> 01:36:31,741
reinvent those technologies so
that they can actually be

1668
01:36:31,741 --> 01:36:34,000
economical in the United States.

1669
01:36:34,000 --> 01:36:35,000
Audrow Nash: Totally.

1670
01:36:35,000 --> 01:36:39,039
Edward Mehr: and I think it's
interesting. I think the first

1671
01:36:39,039 --> 01:36:43,348
cap is easy to argue for. It's,
I think to me, it's a lazy idea.

1672
01:36:43,348 --> 01:36:47,589
and there's a lot of tractions
for it, where we're just saying,

1673
01:36:47,589 --> 01:36:51,427
let's, do what we did in the
1940s. and 1930s. Let's just

1674
01:36:51,427 --> 01:36:55,399
scale it. But there's a reason
we lost it. there's a reason

1675
01:36:55,399 --> 01:36:59,573
that as the, wages and standard
living increased in the United

1676
01:36:59,573 --> 01:37:03,276
States, we offshored these
things to other places. That

1677
01:37:03,276 --> 01:37:07,517
reason is not gone, right? So I
think the manufacturer will not

1678
01:37:07,517 --> 01:37:11,085
look like, the future
manufacturer will not look like

1679
01:37:11,085 --> 01:37:15,057
1940s. We're not going to be
able to replicate what happens

1680
01:37:15,057 --> 01:37:16,000
in China here.

1681
01:37:16,000 --> 01:37:17,000
Audrow Nash: And

1682
01:37:16,000 --> 01:37:18,000
Edward Mehr: you want to beat
China,

1683
01:37:18,000 --> 01:37:22,615
Audrow Nash: it doesn't make any
sense here. our populations are,

1684
01:37:22,615 --> 01:37:27,231
we're looking like there's fewer
younger people and they all want

1685
01:37:27,231 --> 01:37:30,000
to do jobs that are not manual.
So it's

1686
01:37:30,000 --> 01:37:33,717
Edward Mehr: But there's a lot
of push. A lot of people are

1687
01:37:33,717 --> 01:37:37,370
pushing towards okay, let's
bring these capabilities back.

1688
01:37:37,370 --> 01:37:41,024
So we'll see if that pans out. I
think jurys still out. If

1689
01:37:41,024 --> 01:37:44,866
that's, in my opinion, that's
not going to happen, or it will

1690
01:37:44,866 --> 01:37:46,000
be a lot of failed

1691
01:37:46,000 --> 01:37:48,000
Audrow Nash: option where we
just try to bring everything

1692
01:37:48,000 --> 01:37:49,000
Edward Mehr: bring back and
they're going to

1693
01:37:49,000 --> 01:37:52,000
Audrow Nash: have the people for
it. I think

1694
01:37:52,000 --> 01:37:55,351
Edward Mehr: but they're going
to push forward. I think

1695
01:37:55,351 --> 01:37:59,311
politicians will push forward. A
lot of people will push forward,

1696
01:37:59,311 --> 01:38:02,722
oh, let's make thousand
factories here that are same old

1697
01:38:02,722 --> 01:38:06,500
technologies, right? I do think
though, We need to be a little

1698
01:38:06,500 --> 01:38:10,216
bit smarter about it and take
the harder route and create new

1699
01:38:10,216 --> 01:38:14,054
manufacturing technologies that
give us a competitive advantage

1700
01:38:14,054 --> 01:38:17,709
versus the traditional that our
opponents are using, because

1701
01:38:17,709 --> 01:38:21,487
that's how we're going to have a
sustained advantage. same cat

1702
01:38:21,487 --> 01:38:25,081
foundry will not work. It's a
more advanced foundry here is

1703
01:38:25,081 --> 01:38:28,858
going to work. That's going to
be robot powered and it's going

1704
01:38:28,858 --> 01:38:32,696
to be, intelligent. Same sheet
forming and stamping in China is

1705
01:38:32,696 --> 01:38:36,534
not going to work here. We need
to figure out how we can create

1706
01:38:36,534 --> 01:38:40,129
technologies that form sheet
metal in a much more different

1707
01:38:40,129 --> 01:38:43,723
way, in a much more agile way.
Gives you a more significant

1708
01:38:43,723 --> 01:38:47,257
advantage over the current
technology, not just scaling it

1709
01:38:47,257 --> 01:38:50,851
with dollars. So I think that's
basically the direction I'm

1710
01:38:50,851 --> 01:38:54,567
predicting. I think both will
happen. One of them will likely

1711
01:38:54,567 --> 01:38:58,406
fail, and it's not going to be
as fruitful. and then the other,

1712
01:38:58,406 --> 01:39:02,000
I think, is the route to go. The
latter is the route to go.

1713
01:39:02,000 --> 01:39:06,888
Audrow Nash: I think with COVID,
and Like maybe lockdowns in

1714
01:39:06,888 --> 01:39:12,021
China or something. The, like a
lot of textiles were not coming

1715
01:39:12,021 --> 01:39:16,746
to the US and so I think we
innovated quite a bit. I heard

1716
01:39:16,746 --> 01:39:21,879
this somewhere, and I am not a
hundred percent sure, but what I

1717
01:39:21,879 --> 01:39:26,930
ha, what I believe I heard was
that the, Textile manufacturing

1718
01:39:26,930 --> 01:39:31,737
had a big renaissance here
because we evolved or we created

1719
01:39:31,737 --> 01:39:35,974
new ways of doing the
manufacturing so we could have

1720
01:39:35,974 --> 01:39:41,106
the stuff we wanted. and if say,
transport lines and if the, if

1721
01:39:41,106 --> 01:39:45,995
it's harder to get stuff from
abroad, Hopefully a lot of the

1722
01:39:45,995 --> 01:39:50,720
things we can figure out how to
make here in the US out of

1723
01:39:50,720 --> 01:39:55,771
ingenuity, like that story about
the fabrics and what you guys

1724
01:39:55,771 --> 01:40:00,415
are doing, making it more agile,
manufacturing, using our

1725
01:40:00,415 --> 01:40:04,000
strengths, which is technology,
I would say.

1726
01:40:04,000 --> 01:40:06,951
Edward Mehr: yes, they're
similar thinking is also

1727
01:40:06,951 --> 01:40:10,433
existent. in military and to
some extent some of it is good

1728
01:40:10,433 --> 01:40:14,151
like for example we were talking
about okay you know what China

1729
01:40:14,151 --> 01:40:17,751
can have higher production rate
why can't we just manufacture

1730
01:40:17,751 --> 01:40:21,233
drones that are maybe lower
quality but the same production

1731
01:40:21,233 --> 01:40:24,538
rate right and and there's no
type of thinking moves you

1732
01:40:24,538 --> 01:40:28,079
toward maybe we need a little
bit of it but moves you toward

1733
01:40:28,079 --> 01:40:31,738
replicate what's happening in
China which i don't think it was

1734
01:40:31,738 --> 01:40:35,279
like at best we're going to
become as good as China and then

1735
01:40:35,279 --> 01:40:38,584
But then we don't have the
demographic, the workforce to

1736
01:40:38,584 --> 01:40:42,361
back it up. So let's focus on it
like a different way of making.

1737
01:40:42,361 --> 01:40:46,197
Maybe we should make drones that
are, have the same agility, same

1738
01:40:46,197 --> 01:40:49,679
production rate, but maybe last
longer. They're better. and

1739
01:40:49,679 --> 01:40:52,866
let's invest in those
technologies. So anyway, I think

1740
01:40:52,866 --> 01:40:56,584
there is an easy way out, which
is like replicating what we had

1741
01:40:56,584 --> 01:40:58,000
in the past, which might

1742
01:40:58,000 --> 01:41:01,804
Audrow Nash: But may not work
for this kind of thing, as

1743
01:41:01,804 --> 01:41:05,608
you're suggesting. Yeah, and
then there's the other way,

1744
01:41:05,608 --> 01:41:09,208
which is make better ways of
doing things. And that's

1745
01:41:09,208 --> 01:41:13,419
probably the best way to do it.
I forget who I was talking to,

1746
01:41:13,419 --> 01:41:17,698
but someone was saying that, if
we wanted to build an iPhone in

1747
01:41:17,698 --> 01:41:21,706
the US, they have and take all
of this with salt, because I

1748
01:41:21,706 --> 01:41:25,985
don't remember the specifics but
it's like there's a whole city

1749
01:41:25,985 --> 01:41:30,128
in China that has like a million
people that all work for the

1750
01:41:30,128 --> 01:41:34,408
iPhone manufacturers and like we
don't have imagine a full city

1751
01:41:34,408 --> 01:41:38,619
that is a hundred percent around
the fabrication of an iPhone.

1752
01:41:38,619 --> 01:41:42,626
Which is bonkers to consider
here. Like I'm in San Antonio,

1753
01:41:42,626 --> 01:41:46,974
Texas and it is a big city, but
we have so many industries here.

1754
01:41:46,974 --> 01:41:51,253
It's not like we don't just all
work on an iPhone for this kind

1755
01:41:51,253 --> 01:41:52,000
of thing or

1756
01:41:52,000 --> 01:41:53,000
Edward Mehr: no, you're
absolutely right. you look at

1757
01:41:53,000 --> 01:41:55,000
Audrow Nash: to live.

1758
01:41:55,000 --> 01:41:58,729
Edward Mehr: it's machine,
operators running these machines

1759
01:41:58,729 --> 01:42:02,584
over and over, again, making
phones and things like that. But

1760
01:42:02,584 --> 01:42:06,565
yeah, I don't think it will work
here. there's a reason, worked

1761
01:42:06,565 --> 01:42:10,420
in 1940s, but I don't think it
worked now. so yes, I'm in, in

1762
01:42:10,420 --> 01:42:12,000
total agreement with you.

1763
01:42:12,000 --> 01:42:15,000
Audrow Nash: Hell yeah.

1764
01:42:15,000 --> 01:42:19,521
Audrow Nash: What advice do you
have? So say someone is in the

1765
01:42:19,521 --> 01:42:23,969
early part of their career and
they want to get more involved

1766
01:42:23,969 --> 01:42:28,417
in the manufacturing space. how
would they do so? they have a

1767
01:42:28,417 --> 01:42:32,719
background in technology and
they want to get more involved

1768
01:42:32,719 --> 01:42:37,385
in the manufacturing space. what
would, what advice could grease

1769
01:42:37,385 --> 01:42:42,052
their path, make it a bit easier
to get started and productively

1770
01:42:42,052 --> 01:42:43,000
contributing?

1771
01:42:43,000 --> 01:42:46,436
Edward Mehr: I think, obviously
there's like millions of ways to

1772
01:42:46,436 --> 01:42:49,497
get involved, so whatever I'm
suggesting is probably like

1773
01:42:49,497 --> 01:42:51,000
something that I anchored on

1774
01:42:51,000 --> 01:42:52,000
Audrow Nash: Yeah, for sure.

1775
01:42:51,000 --> 01:42:54,686
Edward Mehr: more relevant to
me. I think the big, we have a

1776
01:42:54,686 --> 01:42:58,372
lot of smart people working on a
lot of smart things, right?

1777
01:42:58,372 --> 01:43:01,936
folks in AI and software,
electrical engineering, a lot of

1778
01:43:01,936 --> 01:43:05,253
folks doing a lot of smart
things. I have noticed that

1779
01:43:05,253 --> 01:43:08,694
people think of manufacturing
out of reach, because as a

1780
01:43:08,694 --> 01:43:12,687
society, we don't grow up around
manufacturing anymore, right? We

1781
01:43:12,687 --> 01:43:16,496
grow up in a lot of technology,
but we don't see manufacturing

1782
01:43:16,496 --> 01:43:20,182
day to day, and to some extent,
that removed that agency and

1783
01:43:20,182 --> 01:43:24,114
that removed that believe that,
I can do this, right? one of the

1784
01:43:24,114 --> 01:43:28,046
things that happened to me when
I went at SpaceX, I used to do a

1785
01:43:28,046 --> 01:43:31,794
lot of manufacturing as a kid,
but when I went out to SpaceX,

1786
01:43:31,794 --> 01:43:35,541
the large scale of operation,
the manufacturing operations we

1787
01:43:35,541 --> 01:43:39,166
were running, gave me and a lot
of other people who were at

1788
01:43:39,166 --> 01:43:43,098
SpaceX a lot of confidence. They
thought, oh, it's totally fine,

1789
01:43:43,098 --> 01:43:46,968
I can do this, right? So what I
would suggest to people who are

1790
01:43:46,968 --> 01:43:50,778
smart and have been working in
other disciplines that are much

1791
01:43:50,778 --> 01:43:54,156
more theoretical, you're a smart
guy, just start making

1792
01:43:54,156 --> 01:43:57,904
something. learn to weld, go
learn to do a carpentry, do some

1793
01:43:57,904 --> 01:44:01,836
kind of a fabrication, and maybe
get involved with people who do

1794
01:44:01,836 --> 01:44:05,461
fabrication. But I think the
biggest portion is once you do

1795
01:44:05,461 --> 01:44:09,208
it, you realize, oh, first of
all, it's really fun. Second of

1796
01:44:09,208 --> 01:44:12,833
all. I can do it. Now I can
apply this skill set that I had

1797
01:44:12,833 --> 01:44:16,519
in robotics to do it as well.
when I was doing sheet shaping

1798
01:44:16,519 --> 01:44:20,513
by hand, I was like, okay, I can
do it sheet by sheet, I can make

1799
01:44:20,513 --> 01:44:24,445
a door for a car, whatever. But
then I was like, what, I spend a

1800
01:44:24,445 --> 01:44:27,885
lot of time in robotic and
computer engineering. Maybe I

1801
01:44:27,885 --> 01:44:31,510
can do this. I can, make it,
make, put a robot together and

1802
01:44:31,510 --> 01:44:35,073
build this. I think creativity
comes from when you connect

1803
01:44:35,073 --> 01:44:38,636
multiple disciplines. So go do
those disciplines. Don't be

1804
01:44:38,636 --> 01:44:42,445
scared of going out there and
taking a welding class. Don't be

1805
01:44:42,445 --> 01:44:46,254
scared of go taking a carpentry
class. And then start thinking

1806
01:44:46,254 --> 01:44:49,388
about how you can connect the
dots between multiple

1807
01:44:49,388 --> 01:44:53,197
disciplines that you might be
involved. and I think that's how

1808
01:44:53,197 --> 01:44:56,637
we're going to build those next
generation manufacturing

1809
01:44:56,637 --> 01:45:00,385
technologies that you talked
about and not just replicate the

1810
01:45:00,385 --> 01:45:04,071
past. Is these smart people
coming from world of physics and

1811
01:45:04,071 --> 01:45:07,634
electrical engineering and
robotics and AI doing these old

1812
01:45:07,634 --> 01:45:11,320
mundane tasks that sound like
mundane, but then you're like,

1813
01:45:11,320 --> 01:45:15,007
Oh no, like I can significantly
drive optimizations here and

1814
01:45:15,007 --> 01:45:19,000
improve these techniques by 10X,
20X, 30X. does that, make sense?

1815
01:45:19,000 --> 01:45:22,775
Audrow Nash: Totally get
involved, start building stuff,

1816
01:45:22,775 --> 01:45:26,551
go do a welding class, do a
carpenter class. I think I'm

1817
01:45:26,551 --> 01:45:30,528
gonna follow your advice with
that. 'cause I am, coding all

1818
01:45:30,528 --> 01:45:34,506
day and I love it. But I also
would love to do more with my

1819
01:45:34,506 --> 01:45:38,551
hands, and building stuff. I'm
like, I, have a pretty decent

1820
01:45:38,551 --> 01:45:42,393
background of building things
around, but I don't do much

1821
01:45:42,393 --> 01:45:46,303
lately and I feel like I'm
losing it for this, like simple

1822
01:45:46,303 --> 01:45:49,000
robotics projects have been
challenging.

1823
01:45:49,000 --> 01:45:52,875
Edward Mehr: Yeah. And I think
the good thing about it is it's

1824
01:45:52,875 --> 01:45:56,813
funny, even when you talk about
the craftsmen, you start having

1825
01:45:56,813 --> 01:46:00,813
these conversations of material
and properties. It doesn't sound

1826
01:46:00,813 --> 01:46:04,750
like a material scientist, but
it's so much deep knowledge that

1827
01:46:04,750 --> 01:46:08,688
you gather by, by just talking
to physics world, the world, the

1828
01:46:08,688 --> 01:46:12,375
physics of the world becomes
more accessible to you and you

1829
01:46:12,375 --> 01:46:15,813
can reason it with it in
intuitive ways. When you start

1830
01:46:15,813 --> 01:46:19,250
building versus before, it's
just like textbook and Oh,

1831
01:46:19,250 --> 01:46:23,000
elastic versus plastic strain,
stress strain curve. Like you

1832
01:46:23,000 --> 01:46:26,688
really get it. And then you
start like feeling it with your

1833
01:46:26,688 --> 01:46:30,063
hands, what that means. And it
allows you to have very

1834
01:46:30,063 --> 01:46:33,688
intelligent conversations and
develop very complex topics.

1835
01:46:33,688 --> 01:46:37,375
Much simpler, as opposed to
coming from a metallurgy school

1836
01:46:37,375 --> 01:46:41,125
or whatever. Anyway, I think
it's, it's hard to describe it,

1837
01:46:41,125 --> 01:46:44,813
but I think once you start
putting your hands on things and

1838
01:46:44,813 --> 01:46:48,563
building things, somehow you
feel much more empowered, to do

1839
01:46:48,563 --> 01:46:49,000
things.

1840
01:46:49,000 --> 01:46:52,714
Audrow Nash: Hell yeah. Okay,
we'll end there. Great

1841
01:46:52,714 --> 01:46:56,429
conversation, Ed. Really
appreciate you being on the

1842
01:46:56,429 --> 01:46:57,000
podcast.

1843
01:46:57,000 --> 01:47:00,500
Edward Mehr: same, thanks for,
for the engaging conversation

1844
01:47:00,500 --> 01:47:03,767
and, I hope other people's
enjoyed as well. I definitely

1845
01:47:03,767 --> 01:47:04,000
did.

1846
01:47:04,000 --> 01:47:07,818
Audrow Nash: Bye, everyone.
That's it. You made it. What did

1847
01:47:07,818 --> 01:47:11,891
you think? Is Machina Labs onto
something? Do you see more great

1848
01:47:11,891 --> 01:47:15,455
uses for AI in robotics in
manufacturing? Let me know in

1849
01:47:15,455 --> 01:47:18,000
the comments or on X. See you
next time.
