Getting Physical with AI

1

00:00:00,000 --> 00:00:05,279

And it came back with very perfect prediction within the design space, the data that we had. I

2

00:00:05,280 --> 00:00:11,040

was like, wow, that was quite easy. AI has made great progress over the last couple of years, not

3

00:00:11,040 --> 00:00:17,280

just in language models, but also in physical simulations, robotics, automotive and industrial

4

00:00:17,280 --> 00:00:23,240

manufacturing. It's only been really in the last, let's say, six months that I think this space has

5

00:00:23,240 --> 00:00:28,840

gotten really exciting. And you can create foundational physics models that predict

6

00:00:28,840 --> 00:00:34,640

simulation runs without having to do them. In this episode of AI Cloud Essentials, I sit down with

7

00:00:34,640 --> 00:00:40,600

Richard, who leads physical AI or VIF, and we talk about the rapidly evolving innovations in this

8

00:00:40,600 --> 00:00:47,479

field. Sometimes you have to get out of the AI space to make actual progress. Don't miss

9

00:00:47,480 --> 00:00:48,880

this conversation.

10

00:00:58,000 --> 00:01:03,960

Hello everyone! Welcome to season two of the AI Cloud Essentials, a podcast series brought to you

11

00:01:03,960 --> 00:01:10,200

by Corby. Today we are in episode two and I'm super excited because I'm joined by Richard and

12

00:01:10,200 --> 00:01:15,120

we are going to talk about how AI is breaking into the physical world out of the digital

13

00:01:15,160 --> 00:01:20,560

terminal, into the physical world. So welcome, Richard. Uh, let's kick things off with your

14

00:01:20,560 --> 00:01:26,839

journey in this area. How did you get into the physical simulation? What is so exciting about it?

15

00:01:26,880 --> 00:01:33,360

Would love to hear your perspective on it. Thanks. First of all, great to be here. I probably got into

16

00:01:33,360 --> 00:01:40,280

the whole AI for physics and physical systems long before ChatGPT, long before it was one of the

17

00:01:40,280 --> 00:01:45,480

biggest topics to talk about at the world's leading conferences. Some people might say, I've

18

00:01:45,480 --> 00:01:52,240

been too early. I think I just wasn't the right place half a decade too early. So I

19

00:01:52,600 --> 00:01:57,360

started a company right out of university. I was a PhD student at Imperial College in London

20

00:01:57,480 --> 00:02:03,880

Researching how to optimize aircraft engines with Rolls-Royce. I had a little stint at NASA where I

21

00:02:03,880 --> 00:02:09,560

worked on the Mars rocket trying to optimize that, and then did a research project, actually in

22

00:02:09,560 --> 00:02:15,800

Stanford, where I was supposed to model turbulence. And few people know turbulence is one of the most

23

00:02:15,800 --> 00:02:22,360

difficult problems in the world. It's almost impossible to solve, and I couldn't solve it

24

00:02:22,360 --> 00:02:28,800

either. Big surprise. And so I cheated. I used AI. This was in 2016. It was a big conference in

25

00:02:28,800 --> 00:02:33,400

Stanford. The world's smartest people were there. Nobody could figure out these equations. And I was

26

00:02:33,400 --> 00:02:38,640

like, you know what I'm going to do? I'm gonna take a little bit of data from some real wind tunnel.

27

00:02:38,640 --> 00:02:43,360

I'm going to fit a deep neural network on this, and I'm going to see what this predicts. And it

28

00:02:43,360 --> 00:02:48,839

came back with very perfect prediction within the design space, the data that we had. And I was like,

29

00:02:48,840 --> 00:02:53,960

wow, that was quite easy. Like this is one of the hardest problems in the world. And with a couple

30

00:02:53,960 --> 00:02:58,400

of lines of code and a little bit of training, I could suddenly solve it. When I say I. AI could

31

00:02:58,400 --> 00:03:04,120

suddenly solve it. And so that's I was hooked, I got in, I was like amazing. I'm gonna from now on

32

00:03:04,120 --> 00:03:10,880

work on AI for science and engineering. And so I did. Awesome. That was

33

00:03:10,920 --> 00:03:16,800

really your aha moment. Right. So that this is going to be something really transformative. So

34

00:03:16,840 --> 00:03:21,200

I've been reading a lot about some of your successes and proven wins and case specifics. It

35

00:03:21,200 --> 00:03:26,720

would be fantastic to hear, for example, if you could talk about some of the use cases where it

36

00:03:26,720 --> 00:03:32,320

really transformed the aha moment that you just talked about and how usage of those data and how

37

00:03:32,320 --> 00:03:38,439

those physical simulations really solving the problem. So I think it's fair to say that I've

38

00:03:38,480 --> 00:03:43,720

tried everything you can with AI in engineering, literally everything. Right. Like when you, I

39

00:03:43,720 --> 00:03:48,560

started a startup or my company with a main idea I want to do AI in science and AI in science is

40

00:03:48,560 --> 00:03:54,039

everything. Einsteins is aerospace engineering, rocket engineering. It's automotive, it's pharma,

41

00:03:54,039 --> 00:04:00,879

it's robotics. It's, um, trying to figure out new medical equipment. And over the last nine years,

42

00:04:00,880 --> 00:04:07,880

I've literally tried everything. And one of the first things that I was really excited

43

00:04:07,880 --> 00:04:13,039

by is the fact that physical simulations, the tools that large engineering companies use to

44

00:04:13,040 --> 00:04:18,159

avoid prototyping something and trying to understand just how they work in the virtual

45

00:04:18,160 --> 00:04:22,799

world before building something, tend to be very compute intensive, and then tend to take a very

46

00:04:22,800 --> 00:04:27,079

long time. Right at the beginning, we used dynamic graph neural networks and they were very

47

00:04:27,079 --> 00:04:33,359

impressive. You could very quickly impress a VP of engineering that he could run a simulation 10,000

48

00:04:33,399 --> 00:04:38,800

times faster than they could ever do it before. And it was impressive. But the problem was that

49

00:04:38,800 --> 00:04:45,239

the economics of training versus inference didn't quite work out. People needed to train

50

00:04:45,280 --> 00:04:52,239

1500 2000 simulations to essentially predict one simulation run, and then realize that

51

00:04:52,240 --> 00:04:58,279

that didn't quite work because they had changed the design too much out of the learning space, and

52

00:04:58,279 --> 00:05:02,719

so they couldn't actually predict it. So they'd run another 500 simulations. So long story short,

53

00:05:02,760 --> 00:05:08,319

we were just too early because models weren't quite intelligent enough yet to actually make

54

00:05:08,320 --> 00:05:13,959

predictions and save people design iterations. That all changed massively when Transformers were

55

00:05:13,959 --> 00:05:20,519

introduced, and suddenly you could actually train an AI model on ten, 15,

56

00:05:20,560 --> 00:05:26,519

20 fairly complex simulations with huge design spaces and get a prediction for a future design

57

00:05:26,519 --> 00:05:33,200

run that actually helped you make a decision and was, like, miraculously correct without having to

58

00:05:33,239 --> 00:05:39,079

rerun the simulation. And that only happened really recently. It's only been really in the last,

59

00:05:39,120 --> 00:05:44,760

let's say, six months, that I think this space has gotten really exciting. And you can create

60

00:05:44,800 --> 00:05:50,159

foundational physics models that predict simulation runs without having to do them. Yeah,

61

00:05:50,160 --> 00:05:55,359

that's fantastic to hear, Richard. And I think it also helps in kind of taking care of a lot of

62

00:05:55,359 --> 00:06:00,359

historical data, which in the past people couldn't have used it. Is that a fair statement? Because now

63

00:06:00,359 --> 00:06:04,479

when you're working with, you know, transformer models which can process, you know, not just, you

64

00:06:04,480 --> 00:06:09,959

know, multiple simulations, but a lot of, you know, historical data to kind of predict the future

65

00:06:09,959 --> 00:06:15,559

much more accurately. Is that a fair statement? I've always loved the idea of that. I've gone to

66

00:06:15,559 --> 00:06:20,920

all of the big engineering companies and I've asked them, hey, how many chassis simulations do

67

00:06:20,920 --> 00:06:26,279

you actually have? And they're like, oh yeah, I think we've run 150,000 of those in the last six

68

00:06:26,279 --> 00:06:31,920

months alone, and we probably have a million in storage. And then we tried to build AI models that

69

00:06:31,920 --> 00:06:36,679

actually learn from them. And then you very quickly realize how much of a problem data

70

00:06:36,880 --> 00:06:43,760

uniformity actually is, because, yes, they have 150,000 simulations, but 150 of those

71

00:06:43,760 --> 00:06:47,959

were designed by this person who had chose slightly different set up conditions, and they

72

00:06:47,959 --> 00:06:52,760

changed software vendor six months ago. And so that's a completely different silver. And

73

00:06:52,760 --> 00:06:58,319

comparing this is apple and oranges. And so I think in general that is the idea. But in reality,

74

00:06:58,359 --> 00:07:03,399

just like with every historic data problem, it's remarkably hard to go and sort through all of

75

00:07:03,399 --> 00:07:08,839

this data and actually bring it together. And so with the latest transformer models, I don't think

76

00:07:08,839 --> 00:07:14,240

that is that much of an issue anymore, because you can literally come up with an intelligent model

77

00:07:14,240 --> 00:07:20,439

from fairly few designs and simulations, and people have tried to come up with ways to bridge

78

00:07:20,440 --> 00:07:25,319

the gap. But yes, it's been. That's where the real value is. And it has historically been very, very

79

00:07:25,319 --> 00:07:30,680

hard to unify all of this. Yeah, that's a fantastic insight for our viewers today because, you know, we

80

00:07:30,720 --> 00:07:35,519

kind of assume that, you know, you have tons of data and you can make progress. So now let's pivot

81

00:07:35,520 --> 00:07:39,559

to a little bit more as to because there's a lot of excitement in the industry about where the

82

00:07:39,559 --> 00:07:44,159

future is headed. The potential that we are talking about is not just limited to the physical

83

00:07:44,160 --> 00:07:49,400

simulations, but a lot of other industries. So I would love to get your take on what are the other

84

00:07:49,400 --> 00:07:55,359

industries, where are we actually venturing and some of the hard happening items around physical

85

00:07:55,360 --> 00:08:01,799

robots? Uh, I would love to hear what you guys are up to. And what are you seeing in the industry? So

86

00:08:01,839 --> 00:08:06,200

monolith and I think we took a very different route to all of the other startups and even big

87

00:08:06,200 --> 00:08:11,040

engineering companies in this industry, because we were too early for the simulation space. We

88

00:08:11,040 --> 00:08:17,319

basically gave up on simulations and instead focused on testing, which very different idea. You

89

00:08:17,319 --> 00:08:23,320

basically say, hey, look, everyone, we've tried to simulate what happens with this chassis or in

90

00:08:23,320 --> 00:08:28,359

this crash test, it didn't work. The reality was more complex than any simulation that we can

91

00:08:28,359 --> 00:08:33,039

really run. So we're just going to go and have to test this anyway. And so we came up with this

92

00:08:33,039 --> 00:08:39,400

counterintuitive notion that simulations are good, but they will never be as good as the real

93

00:08:39,400 --> 00:08:45,679

world. And so instead of running simulations, why don't we directly include AI into the real world?

94

00:08:45,679 --> 00:08:52,479

Learn from real physical data, real physical prototypes. And this is where we actually started

95

00:08:52,480 --> 00:08:58,840

getting into robotics, because within robotics, one of the big topics out there is what people call

96

00:08:59,080 --> 00:09:05,799

um, the sim to real gap, where everyone is excited about how I can now run a million simulations of

97

00:09:05,799 --> 00:09:12,640

a robot on a cloud. But I've been arguing for a while, um, that this doesn't really help

98

00:09:12,640 --> 00:09:19,400

you if you ask a robot arm to essentially pick up a plastic bottle in virtual reality, this plastic

99

00:09:19,400 --> 00:09:25,560

bottle is solid. It's firm, it doesn't move. It's rigid because simulating how plastic crumbles

100

00:09:25,560 --> 00:09:30,520

when you touch it is remarkably hard and takes a huge amount of computational effort. That you

101

00:09:30,520 --> 00:09:35,840

could run in two seconds is impossible. And so instead of doing this in virtual reality,

102

00:09:35,879 --> 00:09:40,840

basically you could do it in the real physical world. You could set up a robot, put it into a lab,

103

00:09:40,840 --> 00:09:47,199

and just practice with a water bottle and just see what does it really do, like learning faster

104

00:09:47,200 --> 00:09:53,759

with fewer data points from real physical data is not the same thing as the physical

105

00:09:53,760 --> 00:09:57,799

simulation world, but as a different area of machine learning where you're literally trying

106

00:09:57,840 --> 00:10:04,319

within control systems to just see how do I very quickly learn in real time, in a feedback

107

00:10:04,320 --> 00:10:10,079

loop, in a real world environment? There's a couple of use cases that we've published as monoliths,

108

00:10:10,080 --> 00:10:14,399

where these things have actually yielded great results. There's a case study published with

109

00:10:14,400 --> 00:10:19,559

Nissan where we sort them, got them off the path of, hey, let's try and simulate everything and just

110

00:10:19,559 --> 00:10:24,999

say, look, guys, like in physical testing, there are very simple components in your systems, just like

111

00:10:24,999 --> 00:10:30,479

simple bolts, right? Like you bolt together two metal components should be trivial, but turns out

112

00:10:30,479 --> 00:10:35,639

is surprisingly hard. But somewhere out there is specialists who spend 90 years trying to figure

113

00:10:35,640 --> 00:10:40,880

out how to attach two metal plates with a bolt. And there's a lot of knowledge in this. There's a

114

00:10:40,880 --> 00:10:45,719

lot of hardware testing in this. I mean, okay, why don't we just take all of this hardware data, put

115

00:10:45,719 --> 00:10:51,118

it into a machine learning algorithm and have it learn. If I'm going to go and build this, what's

116

00:10:51,119 --> 00:10:55,680

your recommendation? What's going to happen in the real world? No simulations, just learning from real,

117

00:10:55,719 --> 00:11:00,400

real world experience. And it turned out that worked really well. Like once Nissan started

118

00:11:00,400 --> 00:11:05,879

rolling this out for physical tests, they could reduce testing across all of their chassis by

119

00:11:05,920 --> 00:11:11,478

like 17%, which doesn't sound like a lot. But a lot of these tests are very critical for safety, so

120

00:11:11,520 --> 00:11:17,560

they couldn't be omitted. Awesome. Awesome. So as we are kind of expanding into the other industries

121

00:11:17,560 --> 00:11:22,479

and looking into, could you give us some perspective on what other industry segments are

122

00:11:22,479 --> 00:11:28,199

you kind of expanding into or looking into? Uh, you know, there's a lot of buzz in the industry about

123

00:11:28,199 --> 00:11:34,919

collaborative robots or humanoid robots. Are you looking into those areas as well? So for me, within

124

00:11:34,919 --> 00:11:41,799

physical AI, the biggest areas are, number one, automotive autonomous cars

125

00:11:41,840 --> 00:11:48,760

have made a huge leap in the last three months alone, in particular through releases like Nvidia

126

00:11:48,800 --> 00:11:53,959

Alpa Mayo, where your car cannot argue with you and say, no, I'm not going to go and turn left

127

00:11:53,960 --> 00:12:00,639

because there's a cyclist on the road there and it stops being the it stops. It's not that these

128

00:12:00,679 --> 00:12:05,120

AI models were black boxes. It's like, well, I don't trust my autonomous car because it's a black box.

129

00:12:05,120 --> 00:12:10,119

I don't know what it does. I don't trust it. They cannot argue back and explain in words why

130

00:12:10,119 --> 00:12:15,999

they've tried to do it in a specific way, and even involve you in the process. So that and the fact

131

00:12:16,000 --> 00:12:21,559

that the enormous amount of compute that is now available, thanks to the AI boom, they can train

132

00:12:21,560 --> 00:12:27,119

autonomous cars in a completely different way. So automotive autonomous cars, I think, are the number

133

00:12:27,119 --> 00:12:33,119

one area for me where I'm expecting big jumps. The second area that I think is really huge is

134

00:12:33,120 --> 00:12:39,479

manufacturing. Like I keep saying, like when Jensen Huang goes on stage and says it's going to be a

135

00:12:39,479 --> 00:12:46,199

$50 trillion industry by 2050 physical AI. I keep joking that out of that $50 trillion,

136

00:12:46,240 --> 00:12:53,159

probably 49 trillion is in good old fact, good old fashioned manufacturing.

137

00:12:53,600 --> 00:12:58,679

Like, how do I create a Starbucks cup? Like somebody needs to manufacture that. Somebody needs

138

00:12:58,680 --> 00:13:02,799

to check that the quality of this cup is actually all right. Right. Like there shouldn't be a leak in

139

00:13:02,799 --> 00:13:07,679

that ideally. And you should be able to close the lid on it. That's a lot of manufacturing quality

140

00:13:07,680 --> 00:13:12,479

control. That is a lot of robots and cameras trying to figure out does this actually work?

141

00:13:12,559 --> 00:13:17,439

Nobody wants to look at a Starbucks cups every morning to see whether it's perfect. And so that

142

00:13:17,440 --> 00:13:23,039

is I think, the biggest use case out there. Yeah, I'm mentioning these two first because I think

143

00:13:23,040 --> 00:13:28,279

these are the real big ones. Um, humanoid robots. I know they get the most attention in the media

144

00:13:28,320 --> 00:13:34,999

because like a dancing robot is always fun to watch. But that said, there's just

145

00:13:35,000 --> 00:13:41,839

still a lot of technical challenges that we have to overcome in humanoid robots. It's true that if

146

00:13:41,840 --> 00:13:47,559

you put ChatGPT into a robot, there's this moment where you amaze. Because suddenly this thing can

147

00:13:47,559 --> 00:13:54,159

see. It can talk to you, it can reason and argue back. But the problem is, it still does these

148

00:13:54,159 --> 00:13:59,319

things fairly slowly. There's a little bit of a lag here, right? And it still needs to learn a lot

149

00:13:59,320 --> 00:14:05,799

of different things. And it can very quickly be out of scope when you realize are actually this

150

00:14:05,799 --> 00:14:10,519

is a fantastic robot. And it's learn how to cook, but it doesn't work in my kitchen. I've heard that

151

00:14:10,520 --> 00:14:15,599

right. Like there's problems with renting Airbnbs in California these days because robotic

152

00:14:15,599 --> 00:14:20,399

companies are renting so many kitchens to train their cooking robots that you can't get an Airbnb

153

00:14:20,400 --> 00:14:25,959

in California anymore. It's probably exaggerated, but I think it's one of these funny anecdotes

154

00:14:25,960 --> 00:14:32,798

that just say that it's still quite early within robotics. That said, there have been

155

00:14:32,799 --> 00:14:38,759

moments recently where I'm thinking we might be on the verge of a breakthrough. So the journey has

156

00:14:38,799 --> 00:14:44,159

begun. And, you know, we have a long ways to go. But this is fascinating, Richard. We can have a

157

00:14:44,160 --> 00:14:50,759

separate episode. Just going deeper, dive on this. Um, for the interest of for the sake of time, uh,

158

00:14:50,759 --> 00:14:55,919

I'd love to kind of wrap this up right now and come up with, uh, you know, your parting advice,

159

00:14:55,920 --> 00:15:01,800

share your parting advice to the viewers. Uh, if they're listening to you, what's next that they

160

00:15:01,800 --> 00:15:07,559

should do after this? So I think the most interesting thing I've learned working in AI over

161

00:15:07,560 --> 00:15:13,119

the last eight years is that

162

00:15:13,239 --> 00:15:20,239

sometimes you have to get out of the AI space to make actual progress. Yeah, I know

163

00:15:20,240 --> 00:15:24,999

it's a little bit counter-intuitive, but if you look at what engineering companies make the

164

00:15:24,999 --> 00:15:31,918

fastest progress, like SpaceX, they actually do this in a very simplistic manner. They're

165

00:15:31,919 --> 00:15:36,519

like, let's go build this rocket, let's go test this rocket and see what happens, and then learn

166

00:15:36,519 --> 00:15:41,959

from this and iterate from this. I think engineers who love. Next generation tools, whether that is

167

00:15:41,959 --> 00:15:48,839

advanced simulation or AI, can get very swept up in, oh, I'm going to go and solve this with AI and

168

00:15:48,840 --> 00:15:53,760

then don't realize that in many cases, sometimes there can actually be a very simple and intuitive

169

00:15:53,760 --> 00:15:58,678

solution and are practical, right? Like get out of the simulator and just check whether this thing

170

00:15:58,679 --> 00:16:04,520

drops when you sort of stop holding it, right. Like can sometimes actually solve a lot of problems.

171

00:16:04,559 --> 00:16:10,439

Awesome. Awesome. So my closing thing would be just AI isn't replacing an engineer, it's giving them

172

00:16:10,439 --> 00:16:15,759

the superpowers to use it constructively and learn from it, test more and iterate more. Thank

173

00:16:15,759 --> 00:16:20,799

you everyone for joining us for this episode. We had a fantastic conversation with Richard. I'm

174

00:16:20,800 --> 00:16:25,279

sure you all will love listening to this and stay tuned for more. Thank you.

‍