1
00:00:00,280 --> 00:00:02,400
Hi, and welcome to the Neil 
 
Ashton Podcast. 

2
00:00:03,080 --> 00:00:05,920
In each episode, we explained 
 
some of the fascinating ways 

3
00:00:05,920 --> 00:00:09,080
that science and engineering are

 changing the world around us. 

4
00:00:09,800 --> 00:00:12,760
We talked to leading engineers 

from elite level sports like 

5
00:00:12,840 --> 00:00:16,920
cycling and Formula One to some 
 of the world's top academics to

6
00:00:16,920 --> 00:00:20,490
understand how fluid dynamics, 

machine learning and 

7
00:00:20,490 --> 00:00:23,000
supercomputing are bringing in a
new era of 
 discovery. 

8
00:00:23,960 --> 00:00:27,080
We also hear some of their life 
 stories, their career advice, 

9
00:00:27,640 --> 00:00:30,080
the lessons they've learned on 

the way that I hope will be 

10
00:00:30,080 --> 00:00:33,840
helpful to you too. 
 
So sit back and enjoy this 

11
00:00:33,840 --> 00:00:41,880
episode. 
 
Hi, and welcome back to the Neil

12
00:00:41,880 --> 00:00:44,680
Ashton Podcast. 
 
So today I have a very special 

13
00:00:44,680 --> 00:00:49,850
guest in Prith Banerjee, who is 
 the CTO of Ansys, somebody who 

14
00:00:49,850 --> 00:00:56,360
I am honoured to have on the 
 
podcast because he truly is at 

15
00:00:56,360 --> 00:01:00,560
the position with the knowledge 
 to answer many of the questions

16
00:01:00,560 --> 00:01:03,840
that I have been wondering 
 
myself, but also asking many of 

17
00:01:03,840 --> 00:01:08,280
the guests on this podcast. 
 
So to ask the CTO of one of the 

18
00:01:08,280 --> 00:01:12,680
largest and most important 
 CAE
companies in the world was a 

19
00:01:12,680 --> 00:01:15,600
great honour. 
 
And I hope it, it's good for you

20
00:01:15,600 --> 00:01:18,800
as well to actually hear from, 

you know, the person really at 

21
00:01:18,800 --> 00:01:20,360
the top of one of these big 
 
companies. 

22
00:01:21,040 --> 00:01:22,620
And he's an amazing individual. 
 

23
00:01:22,628 --> 00:01:26,100
Actually, I, I watched some 
videos of interviews with him 
 

24
00:01:26,108 --> 00:01:31,828
over the past few months and I 
was so impressed by his 
 

25
00:01:31,836 --> 00:01:36,688
understanding of these emerging 
areas, but also the way that he 

26
00:01:36,688 --> 00:01:40,018
 was able to explain it in such 
a simple way. 
 

27
00:01:40,026 --> 00:01:44,440
And you'll see him do this in 
the interview today that 
 

28
00:01:44,448 --> 00:01:47,940
really shows that professor in 
him. 
 

29
00:01:47,948 --> 00:01:51,272
And actually, let's talk about 
what his background is. 
 

30
00:01:51,280 --> 00:01:54,610
Well, he was a professor for 
more than 20 years, publishing 


31
00:01:54,618 --> 00:01:58,047
more than 350 papers, 
supervising, you know, nearly 40

32
00:01:58,047 --> 00:02:00,426

 students. 
So really had an amazing career 

33
00:02:00,426 --> 00:02:05,000
 on its own as a professor in 
Illinois, but then went off to 


34
00:02:05,008 --> 00:02:08,360
the startup world. 
And we discussed a lot about 
 

35
00:02:08,368 --> 00:02:11,340
this need for people to 
sometimes go from academia to 
 

36
00:02:11,348 --> 00:02:14,595
startups to, you know, fully 
exploit the ideas they have. 
 

37
00:02:14,603 --> 00:02:18,780
But then he went into the 
corporate world and became, you 

38
00:02:18,780 --> 00:02:22,536
 know, CTO of companies like 
ABB, HP Labs, Schneider 

39
00:02:22,536 --> 00:02:26,560
Electric, 
 and now Ansys for 
the past six years. 
 

40
00:02:26,568 --> 00:02:30,264
What an incredible individual to
have gone through those three 
 

41
00:02:30,272 --> 00:02:34,057
sort of main stages, I guess, 
of, of, of the world that you 
 

42
00:02:34,065 --> 00:02:37,156
could be in, you know, academia,
startups and and industry. 
 

43
00:02:37,164 --> 00:02:40,220
It's amazing because it's also 
one of those questions I've 
 

44
00:02:40,228 --> 00:02:43,716
often asked people on the show, 
you know what, what do you think

45
00:02:43,716 --> 00:02:46,119

 about the differences? 
So here's somebody who's, you 
 

46
00:02:46,127 --> 00:02:50,690
know, done it all and I really 
wanted to ask him some of the 
 

47
00:02:50,698 --> 00:02:53,805
topics that I personally have 
found interesting at, but I 
 

48
00:02:53,813 --> 00:02:58,020
think the community at large who
are into fluid dynamics and CFD 

49
00:02:58,020 --> 00:03:02,958
 and HPC and AI are wondering. 
So I, I put it to him as the CTO

50
00:03:02,958 --> 00:03:05,128

 of one of the biggest 
companies in the world. 
 

51
00:03:05,136 --> 00:03:08,596
So we had a really deep and I 
thought interesting discussion 

52
00:03:08,596 --> 00:03:11,244
about 
 the role of machine 
learning and artificial 

53
00:03:11,244 --> 00:03:14,160
intelligence in CAE. 
 
We've already dived into some of

54
00:03:14,160 --> 00:03:17,560
the details, discussed at 
 
length about foundational 

55
00:03:17,560 --> 00:03:19,400
models. 
 
He came out with some really 

56
00:03:19,400 --> 00:03:23,680
interesting stuff and, and the 

honesty that he had as CTO to 

57
00:03:23,680 --> 00:03:26,560
explain to his board that this 

really is an important thing 

58
00:03:26,560 --> 00:03:29,840
that could even see the end of 

the simulation market as we know

59
00:03:29,840 --> 00:03:31,375
if they don't fully embrace it. 
 

60
00:03:31,383 --> 00:03:34,960
So we talked a lot about that. 
We talked about quantum 
 

61
00:03:34,968 --> 00:03:39,215
computing, how that could be a 
sign of things to come, some 
 

62
00:03:39,223 --> 00:03:41,332
changes which Ansys have been 
working on. 
 

63
00:03:41,340 --> 00:03:46,085
We touched on, you know, HPC, 
GPUs, but we also talked a 
 lot

64
00:03:46,085 --> 00:03:49,710
about the role of startups, the 
roles of industry, what 
 

65
00:03:49,718 --> 00:03:54,547
startups should be trying to do.
And we talked some advice for 
 

66
00:03:54,555 --> 00:03:57,410
students, mid-career and 
everybody about, you know, what 

67
00:03:57,410 --> 00:04:01,775
 they could do to maybe come up 
with the next amazing invention.

68
00:04:01,775 --> 00:04:03,640

 
We touched on open-source, 

69
00:04:03,640 --> 00:04:07,000
closed-source and how we need to

 work with academia. 

70
00:04:07,960 --> 00:04:13,800
And then, you know, we finally 

ended on some advice, I guess, 

71
00:04:13,800 --> 00:04:17,160
to, to, to people and, and 
 
really finished on what he is 

72
00:04:17,160 --> 00:04:19,959
quite an inspiring individual. 

You know, he's written a book. 

73
00:04:19,959 --> 00:04:22,720
It's really amazing, The 
 
Innovation Factory. 

74
00:04:22,800 --> 00:04:24,960
I can put the link on 
 YouTube 
if you're watching it. 

75
00:04:25,640 --> 00:04:29,520
And on that note, you know, if 

you enjoy this, I really would 

76
00:04:29,520 --> 00:04:33,080
appreciate it if you did, you 
 
know, like it, subscribe it. 

77
00:04:33,080 --> 00:04:36,000
The algorithms work that way. 
 
If you if you like it but don't 

78
00:04:36,000 --> 00:04:38,240
interact, unfortunately, that 
 
makes it harder for others to 

79
00:04:38,240 --> 00:04:40,240
find it. 
 
So I don't often say this, but 

80
00:04:40,240 --> 00:04:43,680
I'll I'll say it once every few 
 episodes just because it would 

81
00:04:43,680 --> 00:04:46,000
help. 
 
And also if you're watching this

82
00:04:46,000 --> 00:04:48,680
on YouTube right now, just to 
 
let you know, this is actually 

83
00:04:48,680 --> 00:04:52,640
also available in audio only on 
 Spotify and Apple and vice 

84
00:04:52,640 --> 00:04:54,080
versa. 
 
If you're listening to this and 

85
00:04:54,080 --> 00:04:57,520
you weren't aware, there is also

 a video version on YouTube. 

86
00:04:57,520 --> 00:05:02,120
So yeah, I, I, I really was so 

pleased that he was willing to 

87
00:05:02,120 --> 00:05:03,960
speak. 
 
I found this conversation so 

88
00:05:03,960 --> 00:05:06,200
interesting, and I hope you do 

too. 

89
00:05:06,200 --> 00:05:09,720
So sit back and listen to this 

interview with Prith Banerjee. 

90
00:05:10,160 --> 00:05:13,640
What was your journey to being 

the CTO of one of the most 

91
00:05:13,640 --> 00:05:16,840
important and biggest simulation

 companies in the world? 

92
00:05:16,840 --> 00:05:19,080
How did you how did you get 
 
there? 

93
00:05:19,080 --> 00:05:21,640
I'm sure others would love to 
 
have your position and your job.

94
00:05:21,640 --> 00:05:24,240
So could you tell me a little 
 
bit more about your career and 

95
00:05:24,240 --> 00:05:26,000
how you got to where you are 
 
today? 

96
00:05:26,480 --> 00:05:29,415
So, so Neil, first of all, thank

 you very much for inviting me 

97
00:05:29,415 --> 00:05:31,960
to this. 
 
So I started my career in 

98
00:05:31,960 --> 00:05:34,880
academia. 
 
I have got my PhD in electrical 

99
00:05:34,880 --> 00:05:37,080
and computer engineering from 
 
the University of Illinois 

100
00:05:37,200 --> 00:05:41,560
Urbana-Champaign, and I started 
 as a professor at Urbana. 

101
00:05:41,560 --> 00:05:45,920
I spent the first dozen years 
 
going through the ranks becoming

102
00:05:45,920 --> 00:05:51,320
a full professor, and I was the 
 founding director of 

103
00:05:51,960 --> 00:05:55,920
Computational Science and 
 
Engineering Program at UIUC. 

104
00:05:55,920 --> 00:05:58,760
Illinois has the National 
 
Center for Supercomputing 

105
00:05:58,760 --> 00:06:01,680
Applications, a big place for 
 
HPC. 

106
00:06:02,560 --> 00:06:05,880
And I used to do and my research

 was on developing parallel 

107
00:06:05,880 --> 00:06:07,640
algorithms and parallel 
 
compilers. 

108
00:06:07,640 --> 00:06:10,920
So I've always been working in 

the HPC area. 

109
00:06:12,160 --> 00:06:15,280
So my last two years at 
 
Illinois, I was a founding 

110
00:06:15,280 --> 00:06:17,760
director of computational 
 
science and engineering, which 

111
00:06:17,760 --> 00:06:21,800
is the field of computing of 
 
high performance computing using

112
00:06:22,080 --> 00:06:25,160
HPC to drive sort of science and

 engineering. 

113
00:06:25,160 --> 00:06:27,480
So computational physics, 
 
computational chemistry, 

114
00:06:27,480 --> 00:06:29,760
computational electromagnetics, 
 all of those things. 

115
00:06:29,800 --> 00:06:34,743
And as it turns out, 30 years 
 
later, I have landed up in this 

116
00:06:34,743 --> 00:06:35,840
job. 
 
So that's sort of the 

117
00:06:35,840 --> 00:06:38,440
connection. 
 
And then after Illinois, I went 

118
00:06:38,440 --> 00:06:40,600
to Northwestern. 
 
I was then at the University of 

119
00:06:40,600 --> 00:06:43,120
Illinois Chicago. 
 
So hardcore academic for about 

120
00:06:43,120 --> 00:06:47,880
20 plus years. 
 
After that, I made a hard turn 

121
00:06:47,880 --> 00:06:52,795
into the corporate world. 
 
I was head of HP Labs, and at HP

122
00:06:52,795 --> 00:06:55,840
Labs I used to lead a lot of 
 
work on on high performance 

123
00:06:55,840 --> 00:06:57,680
computing. 
 
We used to build this really 

124
00:06:57,680 --> 00:07:01,680
super-duper high performance 
 
servers, so a lot of cool work 

125
00:07:01,680 --> 00:07:04,600
there. 
 
And then I became CTO at ABB, a 

126
00:07:04,800 --> 00:07:07,800
power and automation company 
 
based in Zurich. 

127
00:07:07,800 --> 00:07:10,760
And then I became CTO at 
 
Schneider Electric and another 

128
00:07:10,760 --> 00:07:12,600
power automation company based 

in France. 

129
00:07:13,120 --> 00:07:19,000
About six and a half years ago I
I joined 
 Ansys as the CTO. 

130
00:07:19,000 --> 00:07:24,560
So this is my third CTO job. 
 
And what Ansys does is we are 

131
00:07:24,560 --> 00:07:28,120
the leading modeling and 
 
simulation company in the world.

132
00:07:28,240 --> 00:07:31,920
We take the world around us, 
 
which is governed by the laws of

133
00:07:31,920 --> 00:07:34,880
physics. 
 
And we take that physics, which 

134
00:07:34,880 --> 00:07:38,640
is explained as second-order 
 
partial differential equations. 

135
00:07:38,640 --> 00:07:42,120
And we solve those physics 
 
through finite element methods, 

136
00:07:42,120 --> 00:07:46,640
finite volume methods using 
 
things like Fluent, which is our

137
00:07:46,640 --> 00:07:49,680
fluid solver, in things like 
 
Mechanical, which is our 

138
00:07:49,680 --> 00:07:52,280
structural code, in things like 
 HFSS, which is an 

139
00:07:52,280 --> 00:07:56,400
electromagnetic solver. 
 
And my role as CTO is to look at

140
00:07:57,000 --> 00:07:59,920
all these amazing products, 
 
what is the future of 

141
00:07:59,920 --> 00:08:02,600
simulation? 
 
What kind of technologies can be

142
00:08:02,600 --> 00:08:09,960
used to drive future products? 

And in my role as CTO, I look at

143
00:08:09,960 --> 00:08:12,560
things like AI, machine 
 
learning, right? 

144
00:08:12,560 --> 00:08:15,080
How does AI/ML improve 
 
simulation? 

145
00:08:15,480 --> 00:08:19,120
HPC, how do you use HPC to 
 
accelerate simulation? 

146
00:08:19,600 --> 00:08:22,320
How, what do you do with sort of

 cloud, right? 

147
00:08:22,320 --> 00:08:24,360
What do you do with platforms or

 digital engineering? 

148
00:08:24,360 --> 00:08:27,640
So that is I have the coolest 
 
job in the company. 

149
00:08:27,640 --> 00:08:30,800
Yeah, looking at the future 
 
future of simulation. 

150
00:08:31,760 --> 00:08:34,600
Yeah, which is why you're 
 
absolutely perfect guest on this

151
00:08:34,600 --> 00:08:37,679
podcast, because your job is 
 
literally to answer, I guess 

152
00:08:37,679 --> 00:08:39,720
some of the questions that that 
 that people have. 

153
00:08:40,000 --> 00:08:45,480
But maybe I love the fact that 

you have had such a great 

154
00:08:45,480 --> 00:08:48,480
academic career and going into 

industry. 

155
00:08:48,480 --> 00:08:50,680
And it's one of the themes I 
 
often ask people, you know, 

156
00:08:50,720 --> 00:08:53,840
academia or industry, what's the

 benefits of both? 

157
00:08:54,240 --> 00:08:59,840
So what do you now, having done 
 both, what do you see as the 

158
00:08:59,840 --> 00:09:02,840
role of academia? 
 
Where, where can academia help, 

159
00:09:02,840 --> 00:09:07,720
let's say, in advancing CAE and 
 where does industry need to do 

160
00:09:07,720 --> 00:09:09,120
it? 
 
And where is the overlap? 

161
00:09:09,400 --> 00:09:12,080
Absolutely. 
 
So so since you're asking a 

162
00:09:12,080 --> 00:09:16,600
career question, I I actually 
 
bypassed one part of my career. 

163
00:09:17,160 --> 00:09:19,680
So I've actually had three 
 
phases in my career. 

164
00:09:19,680 --> 00:09:23,880
I was in academia for 20 years, 
 but in between academia in the 

165
00:09:23,880 --> 00:09:26,200
large corporate world, I was in 
 the startup world. 

166
00:09:26,200 --> 00:09:30,880
I did two start-ups, One was 
 
AccelChip; one was BINACHIP. 

167
00:09:30,880 --> 00:09:34,600
And these were companies started

 out of technologies from the 

168
00:09:34,600 --> 00:09:37,400
university, from one from 
 
Northwestern and one from the 

169
00:09:37,400 --> 00:09:42,280
University of Illinois. 
 
And I did those while in 

170
00:09:42,280 --> 00:09:44,400
universities you can actually go

 on sabbatical. 

171
00:09:44,400 --> 00:09:47,120
So I left, I took leave from the

 university, did my first 

172
00:09:47,120 --> 00:09:50,240
startup, came back to the 
 
university, the second startup 

173
00:09:50,240 --> 00:09:54,640
came back to the university. 
 
So, and literally the reason I 

174
00:09:54,640 --> 00:09:57,760
went from the academic world to 
 the corporate world is because 

175
00:09:57,760 --> 00:10:00,880
of the startups, right? 
 
So in the, So now let me ask you

176
00:10:00,880 --> 00:10:04,760
the question. 
 
In academia, what people do is 

177
00:10:04,760 --> 00:10:07,560
to solve fundamental problems, 

right? 

178
00:10:07,560 --> 00:10:12,440
Really, I mean what I call 
 
Horizon 3 futuristic research 

179
00:10:12,440 --> 00:10:15,000
problems, right? 
 
Where we are trying to really 

180
00:10:15,000 --> 00:10:20,200
understand what is the absolute 
 the fundamentals of of, of 

181
00:10:20,200 --> 00:10:23,600
technology, right And he worked 
 with graduate students and I 

182
00:10:23,600 --> 00:10:27,417
have had in my 20-plus-year 
career 
 right in academia, I 

183
00:10:27,417 --> 00:10:31,052
have had 37 PhD students and 
40-plus master's 
 students with

184
00:10:31,052 --> 00:10:36,760
whom I have published more than 
350 
 technical papers in IEEE 

185
00:10:36,760 --> 00:10:39,840
conference in this and IEEE 
 
Transactions of that and so on 

186
00:10:39,840 --> 00:10:42,040
so forth. 
 
So that's the world of academia 

187
00:10:42,040 --> 00:10:44,880
where you're, you're 
 
researching, you're discovering 

188
00:10:44,880 --> 00:10:49,753
new things and you're publishing

 that work in the latest 

189
00:10:49,753 --> 00:10:51,500
journals and conferences, right?

 

190
00:10:51,508 --> 00:10:55,780
It's all about creating new 
knowledge and then transferring 

191
00:10:55,780 --> 00:10:59,006
 that knowledge to brilliant 
students, right? 
 

192
00:10:59,014 --> 00:11:02,487
So you are educating the 
workforce in the next World, 
 

193
00:11:02,495 --> 00:11:05,290
right? 
So in academia you have two 
 

194
00:11:05,298 --> 00:11:07,667
roles. 
One is invent, create knowledge,

195
00:11:07,667 --> 00:11:09,200
right? 
 
Discover knowledge which you 

196
00:11:09,200 --> 00:11:13,080
publish and then you train 
 
students with the knowledge that

197
00:11:13,080 --> 00:11:15,360
you have created, right? 
 
Train undergraduate students, 

198
00:11:15,360 --> 00:11:19,280
graduate students and so on, 
 
which are the workforces for all

199
00:11:19,280 --> 00:11:22,760
of us, right, in academia and in

 the corporate world to do. 

200
00:11:23,480 --> 00:11:27,037
But what academia does not do is

 we don't build products, 

201
00:11:27,037 --> 00:11:29,365
right? 
And literally, Neil, the reason 

202
00:11:29,365 --> 00:11:34,600
 I did the startups was I was 
frustrated that I was doing all 

203
00:11:34,600 --> 00:11:38,200
 this work, 350 papers, 10-plus 
patents, doing all kinds of 
 

204
00:11:38,208 --> 00:11:39,601
stuff. 
But nobody cared. 
 

205
00:11:39,609 --> 00:11:43,072
Nobody gave a damn right there 
was because it was not showing 


206
00:11:43,080 --> 00:11:47,340
up in any product. 
So when AccelChip was 
 actually

207
00:11:47,340 --> 00:11:51,675
created when I ended a DARPA 
project called the MATCH 
 

208
00:11:51,683 --> 00:11:55,146
compiler and the DARPA PM said, 
great, this is really awesome. 


209
00:11:55,154 --> 00:11:57,152
You should should transfer it to
a company. 
 

210
00:11:57,160 --> 00:11:59,920
So I came to the Bay Area, 
talked to various companies and 

211
00:11:59,920 --> 00:12:02,258
 say, would you like to use this
technology? 
 

212
00:12:02,266 --> 00:12:05,480
They said, absolutely, this 
looks so good, just leave the 

213
00:12:05,480 --> 00:12:08,496
the 
 software copy with us. 
And I looked at them in the eye 

214
00:12:08,496 --> 00:12:11,645
 and said there's no way they 
are going to take this software 

215
00:12:11,645 --> 00:12:15,940
like 
 the only way this really 
commercialise if I were to do it

216
00:12:15,940 --> 00:12:17,650

 myself with my graduate 
students. 
 

217
00:12:17,658 --> 00:12:22,136
So that's kind of why I started 
the first company, AccelChip. 
 

218
00:12:22,144 --> 00:12:27,975
So startups, what they do is 
they actually take a really new 

219
00:12:27,975 --> 00:12:32,302
 idea, something that the world 
has not seen before and get 
 

220
00:12:32,310 --> 00:12:36,622
laser focused on that idea and 
they bring bring that that new 


221
00:12:36,630 --> 00:12:41,702
product to to the market, right.
And I did two of those startups 

222
00:12:41,702 --> 00:12:45,192
 myself and then I came to the 
large corporate world of HP, ABB

223
00:12:45,192 --> 00:12:50,055

 and so on, right? 
But what I have found is the 
 

224
00:12:50,063 --> 00:12:53,958
large companies, they don't have
a single product like Ansys. 
 

225
00:12:53,966 --> 00:12:56,720
We have 70 products, right, in 
simulation, right? 
 

226
00:12:56,728 --> 00:13:00,455
We have Ansys Mechanical, we 
have LS-DYNA, we have Fluent, we

227
00:13:00,455 --> 00:13:03,400

 have this Twin Builder, all 
kinds of products, right? 
 

228
00:13:03,408 --> 00:13:07,180
And the role of a large company 
is to take these products and 
 

229
00:13:07,188 --> 00:13:10,148
evolve their products, right, 
doing continuous innovation. 
 

230
00:13:10,156 --> 00:13:15,010
What features should I have in 
the next release of Fluent, the 

231
00:13:15,010 --> 00:13:17,810
 next release of, of mechanical 
and so on? 
 

232
00:13:17,818 --> 00:13:21,916
So, but the innovation that 
happens in in the corporate 
 

233
00:13:21,924 --> 00:13:24,295
world is more incremental, 
right? 
 

234
00:13:24,303 --> 00:13:27,840
It is what I call Horizon 1. 
I have a product. 
 

235
00:13:27,848 --> 00:13:31,122
So I used to work at HP, right? 
You make computers. 
 

236
00:13:31,130 --> 00:13:33,736
So, so next version of laptop, 
right? 
 

237
00:13:33,744 --> 00:13:36,520
Is an incremental very 
important, but something that 
 

238
00:13:36,528 --> 00:13:40,565
that you need to do. 
We have Ansys Mechanical, 
 it's

239
00:13:40,565 --> 00:13:43,150
a finite element based 
structural solver. 
 

240
00:13:43,158 --> 00:13:45,640
We are doing the next version, 
right? 
 

241
00:13:45,648 --> 00:13:48,307
It's faster, it's a little 
better convergence, better 
 

242
00:13:48,315 --> 00:13:50,840
meshing, but it's still the same
tool, right? 
 

243
00:13:50,848 --> 00:13:55,612
So that's what large companies 
do academia, we invent new 
 

244
00:13:55,620 --> 00:13:58,406
things, right? 
We are doing a hierarchical 
 

245
00:13:58,414 --> 00:14:00,495
octree mesh representation, 
whatever and you 
 publish a 

246
00:14:00,495 --> 00:14:02,280
paper and you're getting a 
patent and so on. 
 

247
00:14:02,288 --> 00:14:06,315
But that is not a product. 
What startups do is take that 
 

248
00:14:06,323 --> 00:14:11,322
work in academia and they 
package it up into this really 


249
00:14:11,330 --> 00:14:14,840
brilliant disruptive innovation,
which I call Horizon 3 
 

250
00:14:14,848 --> 00:14:17,520
innovation, right? 
The truly disruptive innovation 

251
00:14:17,520 --> 00:14:21,145
 always happens in startups. 
Large companies actually 
 

252
00:14:21,153 --> 00:14:24,760
struggle with with, with with 
disruptive innovation. 
 

253
00:14:24,768 --> 00:14:28,056
In fact, Neil, I have done 
broadcasts on this. 
 

254
00:14:28,064 --> 00:14:31,156
I have written a book called The
Innovation Factory, which your 


255
00:14:31,164 --> 00:14:35,255
readers may be interested in. 
And the whole premise of this 
 

256
00:14:35,263 --> 00:14:40,876
book is how does a large company
like ABB or Schneider or or HP 


257
00:14:40,884 --> 00:14:45,042
or or Ansys, the companies that 
have actually worked in the role

258
00:14:45,042 --> 00:14:49,960

 of CTO, right? 
How do these companies try to 
 

259
00:14:49,968 --> 00:14:53,280
foster Horizon 3 disruptive 
innovation, right? 
 

260
00:14:53,288 --> 00:14:57,320
Large companies doing disruptive
innovation and what I say in my 

261
00:14:57,320 --> 00:15:01,010
 book is they they do it through
partnership with academia 
 

262
00:15:01,018 --> 00:15:04,580
because academia is where the 
research in this future 
 

263
00:15:04,588 --> 00:15:08,540
directions is happening and with
startups and bring those so 
 

264
00:15:08,548 --> 00:15:12,503
academia and startups to this 
thing in a concept called open 


265
00:15:12,511 --> 00:15:15,720
innovation. 
And that is what I'm truly 
 

266
00:15:15,728 --> 00:15:18,077
passionate about. 
So I know you asked me a 
 

267
00:15:18,085 --> 00:15:20,260
question about the difference 
between academia and a large 
 

268
00:15:20,268 --> 00:15:24,076
world. 
Academia does discovery of 
 

269
00:15:24,084 --> 00:15:29,138
knowledge Horizon 3, but they 
don't actually make products 
 

270
00:15:29,146 --> 00:15:32,782
this disruptive innovation. 
There are people like me who 
 

271
00:15:32,790 --> 00:15:35,480
leave academia and they build a 
disruptive thing. 
 

272
00:15:35,488 --> 00:15:41,214
But in a concept of a startup, a
startup is laser focused on that

273
00:15:41,214 --> 00:15:44,672

 one product, right? 
The world that has not seen 
 

274
00:15:44,680 --> 00:15:47,310
right, very disruptive, but 
that's the only thing that they 

275
00:15:47,310 --> 00:15:49,214
 do right. 
So they're they're focused on 
 

276
00:15:49,222 --> 00:15:50,815
it. 
And then they'll do the second 


277
00:15:50,823 --> 00:15:53,415
product and the third product. 
Ultimately that will also become

278
00:15:53,415 --> 00:15:57,635

 a large company, at which 
point it will stop doing Horizon

279
00:15:57,635 --> 00:16:00,662
3 
 innovation. 
It will become like the ABBs 
 

280
00:16:00,670 --> 00:16:03,240
of the world right until they 
start. 
 

281
00:16:03,248 --> 00:16:06,707
They then start working with 
with with other other startup 
 

282
00:16:06,715 --> 00:16:07,975
companies. 
Yeah. 
 

283
00:16:07,983 --> 00:16:10,694
I really like how you put that, 
that that's kind of what I was 


284
00:16:10,702 --> 00:16:13,424
getting at. 
And I have to be honest, 
 

285
00:16:13,432 --> 00:16:18,199
particularly coming from Europe,
I think, and that it is slightly

286
00:16:18,199 --> 00:16:21,477

 changing now, there was 
nowhere near the same startup 

287
00:16:21,477 --> 00:16:23,920
culture. 
 
And it's felt like you're an 

288
00:16:23,920 --> 00:16:26,440
academia. 
 
You know, it was almost a dirty 

289
00:16:26,440 --> 00:16:29,280
word to try and commercialise 
 
what you were doing. 

290
00:16:29,600 --> 00:16:31,320
You know, that's not pure 
 
academia. 

291
00:16:31,320 --> 00:16:35,120
You know, you just publish. 
 
And then there was the large 

292
00:16:35,120 --> 00:16:38,560
companies, you know, the Rolls- 
 Royces or whatever of the world

293
00:16:38,560 --> 00:16:41,600
that I remember were funding it,

 But it always felt like the 

294
00:16:41,600 --> 00:16:44,600
technology transfer wasn't the 

same. 

295
00:16:44,880 --> 00:16:47,720
Now, having worked for a US 
 
company and and being more 

296
00:16:47,720 --> 00:16:51,160
exposed to the Bay Area, I'm 
 
kind of seeing how you're right.

297
00:16:51,160 --> 00:16:53,960
This start-ups. 
 
It seems like it's the it's the 

298
00:16:53,960 --> 00:16:59,560
mechanism in between that allows

 these new ideas to to form. 

299
00:16:59,560 --> 00:17:04,040
But what are there any from your

 time now? 

300
00:17:04,118 --> 00:17:07,800
I guess looking at start-ups, 
 
but also having run a startup, 

301
00:17:08,598 --> 00:17:11,720
what sort of general advice 
 
would you give to start-ups? 

302
00:17:11,760 --> 00:17:14,160
I know this is a very difficult 
 question to answer, but you 

303
00:17:14,160 --> 00:17:21,358
know, yeah, like would you, do 

you go in with the mindset that 

304
00:17:21,358 --> 00:17:23,720
someone's going to buy you? 
 
Do you go in the mindset that 

305
00:17:23,720 --> 00:17:26,400
you are going to be the next big

 company? 

306
00:17:26,400 --> 00:17:29,080
You know, how do you think about

 that? 

307
00:17:29,080 --> 00:17:32,000
Or advice you would give maybe 

to start-ups trying to come up 

308
00:17:32,000 --> 00:17:34,360
with new ideas. 
 
The way I would think about a 

309
00:17:34,360 --> 00:17:39,000
startup is if you're doing a 
 
startup just to make money, 

310
00:17:39,640 --> 00:17:41,224
you've got the wrong motivation.

 

311
00:17:41,232 --> 00:17:45,856
The motivation is really you are
trying to solve a problem that 


312
00:17:45,864 --> 00:17:49,840
the world has, right? 
And you see no solution, right? 

313
00:17:49,840 --> 00:17:51,280
 
There's no existing solution 

314
00:17:51,280 --> 00:17:52,680
from the large companies, right?

 

315
00:17:52,688 --> 00:17:56,280
I mean, you're trying to do this
fantastic computer that will 
 

316
00:17:56,288 --> 00:17:58,326
solve the world's problems, 
right? 
 

317
00:17:58,334 --> 00:18:03,500
I mean, and the world doesn't 
have that tool, that solution 
 

318
00:18:03,508 --> 00:18:08,129
today. 
And you have you are maybe half 

319
00:18:08,129 --> 00:18:11,028
 the time the startup founders 
actually come from large 
 

320
00:18:11,036 --> 00:18:14,120
companies, right? 
And they say they see a problem 

321
00:18:14,120 --> 00:18:17,100
 and they say, you know what, 
I'm going to solve this, right? 

322
00:18:17,100 --> 00:18:18,400
 
And typically in a large 

323
00:18:18,400 --> 00:18:22,920
company, the manager will allow 
 you to only work on things that

324
00:18:22,920 --> 00:18:25,960
are incremental, right? 
 
So you have, as I said, you are 

325
00:18:25,960 --> 00:18:27,920
working in HP or making laptops,

 right? 

326
00:18:28,240 --> 00:18:31,859
If you say to HP, I want to 
build 
 a quantum computer, 

327
00:18:31,859 --> 00:18:33,476
right? 
You imagine we say go away, that

328
00:18:33,476 --> 00:18:37,472

 that's not what we do, right? 
So but oftentimes these 
 

329
00:18:37,480 --> 00:18:42,340
problems come out and look at 
you and say this needs to be 
 

330
00:18:42,348 --> 00:18:44,140
solved. 
And you are, you are just, you 


331
00:18:44,148 --> 00:18:46,935
have this burning passion to 
solve that problem. 
 

332
00:18:46,943 --> 00:18:50,960
And you sometimes your company 
manager will allow you to do it 

333
00:18:50,960 --> 00:18:51,820
 right. 
Then you're lucky. 
 

334
00:18:51,828 --> 00:18:54,532
Then the company is actually 
allowing you to do Horizon 3 
 

335
00:18:54,540 --> 00:18:57,512
innovation. 
But 90% of the time you will not

336
00:18:57,512 --> 00:19:00,272

 be able to do it right. 
And then you say, what, what 
 

337
00:19:00,280 --> 00:19:02,795
choice do I have? 
You should then do a startup, 
 

338
00:19:02,803 --> 00:19:05,589
try to follow your passion, 
follow your dreams and do it 
 

339
00:19:05,597 --> 00:19:07,880
right. 
That's how most entrepreneurs 
 

340
00:19:07,888 --> 00:19:13,401
start startups, right? 
The other way is for academia, 


341
00:19:13,409 --> 00:19:16,045
academic people, right? 
And so literally startups come 


342
00:19:16,053 --> 00:19:19,002
from two ends. 
Either it's an academic who has 

343
00:19:19,002 --> 00:19:22,992
 solved a really hard problem 
and say, OK, now we want to 
 

344
00:19:23,000 --> 00:19:26,140
commercialize it like me. 
And again, I am just a very 
 

345
00:19:26,148 --> 00:19:29,318
small person, but there's so 
many more famous people who came

346
00:19:29,318 --> 00:19:31,448

 from academia and some 
absolutely wonderful companies, 

347
00:19:31,448 --> 00:19:33,720
right? 
 
And I mentioned them in my book.

348
00:19:34,160 --> 00:19:37,680
And and then there's this 
 
startup that happened from. 

349
00:19:37,680 --> 00:19:41,320
So I would say 80% of the 
 
startup founders actually come 

350
00:19:41,320 --> 00:19:42,840
from the large corporate world, 
 right? 

351
00:19:42,840 --> 00:19:46,360
And then they have found a 
 
problem solve it and then they 

352
00:19:46,680 --> 00:19:50,960
start one company, they start a 
 second company now with you 

353
00:19:50,960 --> 00:19:54,847
asked a question, what shop does

 ultimately, yes, so you, you 

354
00:19:54,847 --> 00:19:57,080
get motivated by solving the 
world's 
 problems. 

355
00:19:57,080 --> 00:19:59,120
But of course there is a second 
 motivation. 

356
00:19:59,120 --> 00:20:01,440
I I would like to make some 
 
money out of it, right. 

357
00:20:02,520 --> 00:20:07,600
So the way you pick a problem, 

right, you should pick a problem

358
00:20:07,600 --> 00:20:10,462
that has a large market, right? 
 

359
00:20:10,470 --> 00:20:14,085
And so how do you establish the 
market? 
 

360
00:20:14,093 --> 00:20:17,947
That is the hardest thing for a 
startup entrepreneur to do right

361
00:20:17,947 --> 00:20:20,520

 and. 
So oftentimes you say, well, 
 

362
00:20:20,528 --> 00:20:24,482
what's the market for GPUs? 
Well, you can take the look at, 

363
00:20:24,482 --> 00:20:27,150
 look at NVIDIA and and AMD and 
so on. 
 

364
00:20:27,158 --> 00:20:30,216
And it's OK, These are people 
who are making GPUs, they are 
 

365
00:20:30,224 --> 00:20:32,545
selling this many GPUs. 
And so the market for GPU is 
 

366
00:20:32,553 --> 00:20:34,712
this. 
And if you are a new startup and

367
00:20:34,712 --> 00:20:37,534

 you do another GPU, you know 
exactly what that market is, 
 

368
00:20:37,542 --> 00:20:41,007
right? 
What's the market for, for 
 

369
00:20:41,015 --> 00:20:41,615
eyeglasses? 
Eyeglasses. 
 

370
00:20:41,623 --> 00:20:44,255
You look at all the people who 
are wearing eyeglasses. 
 

371
00:20:44,263 --> 00:20:49,062
You can say that, but suppose 
you are a startup you have you 


372
00:20:49,070 --> 00:20:54,216
are inventing a device such as 
blind men can see. 
 

373
00:20:54,224 --> 00:21:00,481
OK, that device does not exist. 
You do a Google search of market

374
00:21:00,481 --> 00:21:04,616

 for device for blind men per 
se is 0 because there is no 

375
00:21:04,616 --> 00:21:06,926
product 
 in that area. 
I mean, I'm just giving an 
 

376
00:21:06,934 --> 00:21:09,160
example. 
Maybe today there is, but there 

377
00:21:09,160 --> 00:21:11,882
 isn't, right? 
So then you say, oh, the market 

378
00:21:11,882 --> 00:21:13,320
 is 0, therefore it's a bad 
idea. 

379
00:21:13,320 --> 00:21:16,280
I should not do it because those

 marketing things done by 

380
00:21:16,280 --> 00:21:18,440
companies like Gartner or 
 
Dataquest, right? 

381
00:21:18,640 --> 00:21:22,240
They are only looking at at 
 
markets where products exist, 

382
00:21:22,240 --> 00:21:23,760
right? 
 
What's the market for the cloud?

383
00:21:24,120 --> 00:21:30,040
It is $100 billion, right? 
 
The market for cloud before Jeff

384
00:21:30,040 --> 00:21:32,578
Bezos invented AWS was 0 right? 
 

385
00:21:32,586 --> 00:21:37,960
So right it it took a person of 
Jeff's imagination says that the

386
00:21:37,960 --> 00:21:41,620

 market is this if I could 
build it, right? 
 

387
00:21:41,628 --> 00:21:46,307
So then for that, that device 
that blind men can see, right? 


388
00:21:46,315 --> 00:21:48,629
I'm the entrepreneur. 
I'm trying to to find the 
 

389
00:21:48,637 --> 00:21:51,318
market. 
I say, well, how many blind men 

390
00:21:51,318 --> 00:21:55,124
 are there in the world, right, 
that I know I have 10 billion 
 

391
00:21:55,132 --> 00:21:58,854
people on the planet. 
I don't know, maybe 3 million 
 

392
00:21:58,862 --> 00:22:02,483
people are blind. 
How much would they pay for it? 

393
00:22:02,483 --> 00:22:04,080
 
Well, I pay, I go to Lens- 

394
00:22:04,080 --> 00:22:06,480
Crafters and buy these glasses 

for $200.00. 

395
00:22:06,840 --> 00:22:10,400
So at least I'm not blind. 
 
But I'm paying something to 

396
00:22:10,400 --> 00:22:14,400
improve my vision. 
 
So my at least I'll pay 200, 

397
00:22:14,400 --> 00:22:18,040
maybe 300. 
 
So 300 times 100 million blind 

398
00:22:18,040 --> 00:22:20,880
people. 
 
That's the $30 billion market. 

399
00:22:21,160 --> 00:22:22,756
That's how you size the market. 
 

400
00:22:22,764 --> 00:22:26,526
So you have a choice of making a
device such that blind men can 


401
00:22:26,534 --> 00:22:29,542
see. 
The market is $30 billion versus

402
00:22:29,542 --> 00:22:34,075
a 
 chair with 9 legs, right? 
And the market for that is only 

403
00:22:34,075 --> 00:22:36,120
 $2.00. 
You should pick the first one, 


404
00:22:36,128 --> 00:22:39,840
even though that is a harder 
problem to work on because if 
 

405
00:22:39,848 --> 00:22:42,720
you're successful, you will 
solve the world's problem. 
 

406
00:22:42,728 --> 00:22:46,822
And it's a large problem versus 
inventing a chair with 9 legs, 


407
00:22:46,830 --> 00:22:50,018
which is a simple thing because 
you know, I have a chair with 
 

408
00:22:50,026 --> 00:22:52,736
four legs. 
It is easy to do with 9 legs, 
 

409
00:22:52,744 --> 00:22:57,742
but the market is only only two.
That's the simplistic way that I

410
00:22:57,742 --> 00:23:00,623

 can I can I can explain the 
world of. 
 

411
00:23:00,631 --> 00:23:03,110
Startups. 
And I see that a little bit with

412
00:23:03,110 --> 00:23:09,410
simulation is that it 
 is 
difficult probably in CFD or CAE

413
00:23:09,410 --> 00:23:14,286
world to really 
 appreciate the
difference I guess between 

414
00:23:14,286 --> 00:23:17,520
theoretical and 
 would anybody 
actually use it? 

415
00:23:18,080 --> 00:23:20,360
You know, like there's a 
 
difference between saying, oh, 

416
00:23:20,360 --> 00:23:26,320
we could make CFD 10 times 
 
faster, but even if it was 10 

417
00:23:26,320 --> 00:23:29,040
times faster, it doesn't mean 
 
everybody's going to pick your 

418
00:23:29,040 --> 00:23:30,760
software because they may not 
 
trust you. 

419
00:23:30,760 --> 00:23:35,328
They may prefer, you know, So I 
 guess this is where it becomes 

420
00:23:35,328 --> 00:23:38,600
even harder, doesn't it? 
 
When you're, you know, the cloud

421
00:23:38,600 --> 00:23:41,880
was such a massive new thing. 
 
It's so clear. 

422
00:23:41,880 --> 00:23:47,040
I guess most start-ups are more 
 are not as revolutionary, you 

423
00:23:47,040 --> 00:23:48,880
know, and they're probably the 

harder ones, aren't they? 

424
00:23:48,880 --> 00:23:54,200
Because there is a value, but 
 
it's sort of harder to to figure

425
00:23:54,200 --> 00:23:56,160
out. 
 
And I guess maybe this leads 

426
00:23:56,160 --> 00:23:58,800
nicely because one of the things

 that a lot of people have seen

427
00:23:59,600 --> 00:24:04,780
is a huge growth now in the 
AI/ML 
 world, you know, 

428
00:24:04,780 --> 00:24:06,710
obviously for large language 
models. 
 

429
00:24:06,718 --> 00:24:11,248
But I think personally, what's 
excited me is seeing how much of

430
00:24:11,248 --> 00:24:17,200

 this is now slowly moving into
the scientific world and the 
 

431
00:24:17,208 --> 00:24:22,840
potential impact it has on 
accelerating traditional, you 
 

432
00:24:22,848 --> 00:24:27,262
know, CAE codes. 
And I know you have your own 
 

433
00:24:27,270 --> 00:24:31,130
product as well, Ansys SimAI, 
but I was just wanting to get 

434
00:24:31,130 --> 00:24:36,057
maybe 
 some of your thoughts on
where you see the use of AI/ML 

435
00:24:36,057 --> 00:24:41,160
today 
 short term and you know,
what's the what's the think big?

436
00:24:41,160 --> 00:24:42,840

 
What's the art of the possible 

437
00:24:43,240 --> 00:24:45,440
that you think this could 
 
become? 

438
00:24:46,320 --> 00:24:48,160
It's great. 
 
That's a great question. 

439
00:24:48,160 --> 00:24:54,508
So let me explain my 
 thought, 
right, just by going in the area

440
00:24:54,508 --> 00:24:56,629
of simulation itself, 
 right. 
So I want to explain the 
 

441
00:24:56,637 --> 00:24:58,620
problem. 
So when you're looking at 
 

442
00:24:58,628 --> 00:25:02,845
simulation of say, a fluids 
problem, right, the problem is 


443
00:25:02,853 --> 00:25:07,120
formulated in the ideal world as
Navier–Stokes equations, right? 

444
00:25:07,120 --> 00:25:09,120
 
You have the, the governing 

445
00:25:09,120 --> 00:25:11,280
equations, you have energy 
 
conservation, so on. 

446
00:25:11,280 --> 00:25:13,360
And those are second-order PDEs,

 right? 

447
00:25:13,960 --> 00:25:18,000
So you can write those PDEs. 
 
And when you went to college, 

448
00:25:18,000 --> 00:25:19,840
right? 
 
You can take a very simple 

449
00:25:20,400 --> 00:25:24,775
differential equation, right? 
 
Linear, whatever the simplest 

450
00:25:24,775 --> 00:25:28,314
one you could analytically 
solve, 
 right, is E to the 

451
00:25:28,314 --> 00:25:31,120
power -2 whatever sum. 
 
This is how the equations go, 

452
00:25:31,120 --> 00:25:33,960
right? 
 
But in the practical world, 

453
00:25:33,960 --> 00:25:37,360
right, these problems have the 

CAD geometries are so 

454
00:25:37,360 --> 00:25:41,280
complicated by the time you take

 the CAD, define the boundary 

455
00:25:41,280 --> 00:25:43,960
conditions and so on, and you 
 
have the Navier–Stokes equations

456
00:25:44,560 --> 00:25:47,760
to solve it, right? 
 
It is impossible to solve it 

457
00:25:48,040 --> 00:25:49,840
analytically. 
 
So you have to solve it 

458
00:25:49,840 --> 00:25:54,120
numerically. 
 
So you take those PDEs and you 

459
00:25:54,120 --> 00:25:57,880
discretise them, right? 
 
So you do say finite elements, 

460
00:25:57,880 --> 00:26:01,640
right? 
 
You take this whatever kind of 

461
00:26:01,640 --> 00:26:04,360
thing and you break it up into 

1000 elements, right? 

462
00:26:04,520 --> 00:26:07,840
And in each, the finite element 
 method says on each element 

463
00:26:07,840 --> 00:26:09,440
those governing equations will 

work. 

464
00:26:09,600 --> 00:26:14,720
So you solve it on that element 
 with the boundary conditions of

465
00:26:14,720 --> 00:26:18,760
the other nodes that are next to

 you and you keep iterating on 

466
00:26:18,760 --> 00:26:20,560
you. 
 
And that's how all our numerical

467
00:26:20,560 --> 00:26:24,040
methods work, right? 
 
The trouble with these numerical

468
00:26:24,040 --> 00:26:28,320
methods is the trade-off become 
 between accuracy and speed, 

469
00:26:28,320 --> 00:26:30,720
right? 
 
So suppose you solve that 

470
00:26:30,720 --> 00:26:37,000
problem, the CFD with whatever, 
 with say 1000 elements, right? 

471
00:26:37,440 --> 00:26:42,200
And you get an accuracy which is

 about 10% error, which may be 

472
00:26:42,200 --> 00:26:44,200
fine for you. 
 
I said yeah, I like it, right? 

473
00:26:44,200 --> 00:26:48,840
And you solve that in an hour. 

Say I don't like 10% error, I 

474
00:26:48,840 --> 00:26:52,760
wanted to be more accurate. 
 
It is very easy in our world to 

475
00:26:52,760 --> 00:26:57,040
just instead of 1000 elements do

 100,000 elements, right? 

476
00:26:57,040 --> 00:27:02,840
You do finer meshes and it will 
 be 1% error, right? 

477
00:27:03,040 --> 00:27:06,600
But then in instead of 1000 
 
hours to run, it will take you 

478
00:27:06,840 --> 00:27:11,000
100,000 hours to run, right? 
 
So the trade-off of accuracy and

479
00:27:11,000 --> 00:27:15,480
speed in our world of CAE 
 
simulation CFD is, is this 

480
00:27:15,480 --> 00:27:17,480
problem right, the accuracy 
 
versus speed. 

481
00:27:18,160 --> 00:27:21,560
And we want both. 
 
We want both accuracy and speed.

482
00:27:22,840 --> 00:27:27,600
And then furthermore, the third 
 thing is these things are so 

483
00:27:28,000 --> 00:27:29,800
complicated in terms of 
 
convergence. 

484
00:27:29,800 --> 00:27:33,400
Sometimes you do these crazy 
 
things with the meshing, it 

485
00:27:33,400 --> 00:27:35,440
doesn't converge. 
 
So wow, my God, I didn't 

486
00:27:35,720 --> 00:27:38,155
converge. 
 
Oh, it didn't converge because 

487
00:27:38,155 --> 00:27:40,920
of this. 
 
I should use Mosaic meshing. 

488
00:27:40,920 --> 00:27:44,160
I should use tet meshing. 
 
So there are these zillion tools

489
00:27:44,160 --> 00:27:49,360
that I have at my disposal and 

the the CAE analyst is using all

490
00:27:49,360 --> 00:27:52,160
of these things and sometimes it

 works, sometimes it doesn't. 

491
00:27:52,440 --> 00:27:57,840
So it's not that easy to use. 
 
Imagine a tool out there that'll

492
00:27:57,840 --> 00:28:02,920
say, hey, me, run this thing for

 a a external aerodynamics of a

493
00:28:02,920 --> 00:28:06,800
Boeing 777 airplane, right? 
 
You just give it in English and 

494
00:28:07,280 --> 00:28:12,520
automatically it sets the 
 
settings for STAR-CCM+ or or Exa

495
00:28:12,520 --> 00:28:16,040
from Dassault or Fluent. 
 
It just does it like that's the 

496
00:28:16,600 --> 00:28:22,240
ultimate Holy Grail. 
 
So in our world, the problem is 

497
00:28:22,240 --> 00:28:27,040
you have to go be accurate. 
 
You have to be fast, it has to 

498
00:28:27,040 --> 00:28:30,280
be easy to use and converge all 
 the time. 

499
00:28:30,280 --> 00:28:35,200
That is the Holy Grail. 
 
So in my role as CTO, I look at 

500
00:28:35,200 --> 00:28:39,480
all the solvers, I say how can I

 get to that current state to 

501
00:28:39,480 --> 00:28:42,600
make it more accurate, faster, 

easy to use and so on, right. 

502
00:28:42,880 --> 00:28:47,840
So I have, one of the things is 
 I have a pillar on numerical 

503
00:28:47,840 --> 00:28:51,400
methods and we are just with 
 
advanced numerical methods, 

504
00:28:51,400 --> 00:28:54,400
right? 
 
We are doing without using high 

505
00:28:54,400 --> 00:28:56,640
performance computing, without 

using AI/ML. 

506
00:28:56,920 --> 00:29:00,560
I'm trying to make it faster, 
 
accurate, for example doing 

507
00:29:00,560 --> 00:29:03,920
better meshing, for example 
 
using higher order methods, 

508
00:29:03,920 --> 00:29:06,200
right? 
 
How about using hierarchical 

509
00:29:06,200 --> 00:29:08,040
octree? 
 
So it's a sequential algorithm, 

510
00:29:08,040 --> 00:29:12,080
but just using smart things in 

the numerical method itself you 

511
00:29:12,080 --> 00:29:15,400
make it faster, easy to use, 
 
converge all the time and so on.

512
00:29:16,440 --> 00:29:18,862
The second pillar is HPC, right?

 

513
00:29:18,870 --> 00:29:24,152
I mean, again, you worked at AWS
and you have all those high 
 

514
00:29:24,160 --> 00:29:27,624
performance computing, right? 
So we have we, we, we take, we, 

515
00:29:27,624 --> 00:29:32,176
 we take an algorithm and we 
parallelise it, put it on 100 
 

516
00:29:32,184 --> 00:29:35,320
processors using shared memory 
or message passing with 
 

517
00:29:35,328 --> 00:29:38,208
distributed sort of data 
decomposition, all with GPUs. 
 

518
00:29:38,216 --> 00:29:40,752
So there are all these different
things basically. 
 

519
00:29:40,760 --> 00:29:44,525
But this is what I call brute 
force acceleration, right? 
 

520
00:29:44,533 --> 00:29:49,686
I have a job that I have decided
that I will use 1,000,000 
 

521
00:29:49,694 --> 00:29:52,448
elements, right? 
So because of accuracy I have 
 

522
00:29:52,456 --> 00:29:55,052
and it's taking me 1000 hours to
run. 
 

523
00:29:55,060 --> 00:30:00,102
If I had 100 processors, the 
best I can get is get 100-times 

524
00:30:00,102 --> 00:30:02,726
 speedup and run it in 10 hours,
right? 
 

525
00:30:02,734 --> 00:30:06,350
So within that I use shared 
memory message passing GPUs XYZ 

526
00:30:06,350 --> 00:30:09,940
 and now we are looking at 
quantum computing also to speed 

527
00:30:09,940 --> 00:30:13,241
 things up, right? 
But that's what I call brute 
 

528
00:30:13,249 --> 00:30:17,598
force parallelism, right? 
The third pillar that we have is

529
00:30:17,598 --> 00:30:19,650

 AI/ML, which is sort of your 
question. 
 

530
00:30:19,658 --> 00:30:25,263
So AI/ML has been used in a 
variety of fields, but we and it

531
00:30:25,263 --> 00:30:29,050

 has been used in, as you know,
for, for recommendation engines 

532
00:30:29,050 --> 00:30:31,304
 for this and so on. 
Hey, which restaurant should I 


533
00:30:31,312 --> 00:30:33,214
go to? 
It's wonderful for those things,

534
00:30:33,214 --> 00:30:35,140

 right? 
Or ChatGPT allowing you to 
 

535
00:30:35,148 --> 00:30:39,282
write wonderful poetry and text.
But the question that we asked 


536
00:30:39,290 --> 00:30:42,694
is, can AI/ML be applied to 
numerical-method simulation, 
 

537
00:30:42,702 --> 00:30:45,140
right? 
And that's when I joined the 
 

538
00:30:45,148 --> 00:30:47,875
company six years ago, my CEO 
said, what do you want to work 


539
00:30:47,883 --> 00:30:50,544
on? 
I said, I want to work on AI. 
 

540
00:30:50,552 --> 00:30:57,192
And the early work on AI that we
did was to say, OK, let's take a

541
00:30:57,192 --> 00:31:01,894

 black box solver like Fluent 
which is a fluid solver, 
 

542
00:31:01,902 --> 00:31:04,740
right? 
Give it an initial condition, 
 

543
00:31:04,748 --> 00:31:07,690
boundary condition and you get 
the output. 
 

544
00:31:07,698 --> 00:31:11,596
And with this input and output 
you train an AI model, right? 
 

545
00:31:11,604 --> 00:31:14,440
And you see you have this new 
six- stage neural network, 

546
00:31:14,440 --> 00:31:16,120
right? 
 
And you are you have these 

547
00:31:16,120 --> 00:31:17,120
weights of the neural networks. 
 

548
00:31:17,128 --> 00:31:18,595
You don't know what the weights 
are. 
 

549
00:31:18,603 --> 00:31:22,352
So you start with some random 
weights with some random weights

550
00:31:22,352 --> 00:31:27,588

 on the neurons, right? 
You, you here is an input, here 

551
00:31:27,588 --> 00:31:30,582
 is the output. 
So with random weights you will 

552
00:31:30,582 --> 00:31:33,058
 predict an output which will be
completely wrong. 
 

553
00:31:33,066 --> 00:31:36,585
There is an error at the output.
You say now that there is an 
 

554
00:31:36,593 --> 00:31:39,115
error. 
How do I minimize the error? 
 

555
00:31:39,123 --> 00:31:42,425
I do backpropagation to adjust 
the weights of neural networks 


556
00:31:42,433 --> 00:31:46,304
so that my error is zero for 
this input output combination. 


557
00:31:46,312 --> 00:31:51,064
Then I give it a second input 
with a different boundary 
 

558
00:31:51,072 --> 00:31:54,835
condition, different whatever, 
and with now the previous set of

559
00:31:54,835 --> 00:31:58,765

 weights. 
I run it, I get a predicted 
 

560
00:31:58,773 --> 00:32:01,340
output. 
I have a new output from Fluent.

561
00:32:01,340 --> 00:32:02,640

 
Again there is an error. 

562
00:32:02,640 --> 00:32:04,360
I said, oh, I need to fix the 
 
error. 

563
00:32:04,640 --> 00:32:07,960
So I do backpropagation to 
 
change the weights again. 

564
00:32:09,680 --> 00:32:13,960
And then I do the third input 
 
with the first two set of 

565
00:32:13,960 --> 00:32:17,200
weights and my third input. 
 
I keep iterating. 

566
00:32:17,400 --> 00:32:23,680
After about 100–200 cases, I 
 
kind of get, I converge on the 

567
00:32:23,680 --> 00:32:25,680
set of weights on the neural 
 
network, right? 

568
00:32:25,680 --> 00:32:28,240
And within that there's all 
 
kinds of there's choices, right?

569
00:32:28,400 --> 00:32:30,760
Should I have a six-stage 
 
network? 

570
00:32:30,760 --> 00:32:32,240
Should I have a eight-stage 
 
network? 

571
00:32:32,440 --> 00:32:33,600
How many? 
 
What's the depth? 

572
00:32:33,600 --> 00:32:36,880
What's the depth, right? 
 
And that ties to the parameter 

573
00:32:36,880 --> 00:32:39,600
size of your, of your network, 

right? 

574
00:32:39,720 --> 00:32:44,520
But assuming you have done all 

that, right, that's what SimAI 

575
00:32:44,520 --> 00:32:49,680
does. 
 
So SimAI is a platform which 

576
00:32:49,680 --> 00:32:56,880
allows a customer to take their 
 problem their sets of designs. 

577
00:32:57,720 --> 00:33:03,760
Use our tool Fluent for fluid 
 
dynamics or Ansys Mechanical for

578
00:33:03,760 --> 00:33:06,200
structures or HFSS for 
 
electromagnetics. 

579
00:33:06,440 --> 00:33:11,354
And you, Mr. Customer, use SimAI

 platform to train the AI 

580
00:33:11,354 --> 00:33:16,268
models on your problem and then 
train 
 it for the 1st 100 

581
00:33:16,268 --> 00:33:20,714
designs that you have and the 
101st design 
 instead of taking

582
00:33:20,714 --> 00:33:23,632
100 hours, it will run in a 
minute. 
 

583
00:33:23,640 --> 00:33:29,560
That's the value proposition. 
Now the AI is only as good as 
 

584
00:33:29,568 --> 00:33:32,145
the data you train it with, 
right? 
 

585
00:33:32,153 --> 00:33:36,766
So if you train it with this 
picture of you have an SUV, 
 

586
00:33:36,774 --> 00:33:39,580
right? 
You, you train it with this SUV 

587
00:33:39,580 --> 00:33:42,600
 from Toyota, there are 10 
different versions of 
 

588
00:33:42,608 --> 00:33:46,600
Highlander, the, the, the 
4Runner, the RAV4 or 
 whatever.

589
00:33:46,600 --> 00:33:51,080
And then also the SUVs from, 
 
from Hyundai and the SUVs from 

590
00:33:51,280 --> 00:33:53,360
Ford. 
 
So you are training it with 

591
00:33:53,400 --> 00:33:56,760
SUVs, it learns, then you give 

it an airplane. 

592
00:33:57,760 --> 00:34:02,160
I have not seen this before. 
 
And AI is only good as the data 

593
00:34:02,160 --> 00:34:07,040
it has been trained on. 
 
But you may say, oh, therefore 

594
00:34:07,040 --> 00:34:10,320
it's not, not not useful. 
 
It is actually useful because if

595
00:34:10,320 --> 00:34:13,120
you work for a company like 
 
Airbus, right, you're making 

596
00:34:13,120 --> 00:34:15,480
airplanes or Boeing, you're 
 
making airplanes. 

597
00:34:15,800 --> 00:34:19,080
You're not going to go from one 
 airplane to tomorrow doing a 

598
00:34:19,080 --> 00:34:20,840
submarine, right? 
 
So you're actually doing only 

599
00:34:20,840 --> 00:34:23,320
airplanes. 
 
So it actually works. 

600
00:34:23,400 --> 00:34:27,600
There is value in subtle 
 
variations and that's what 

601
00:34:27,600 --> 00:34:29,679
designers do, right? 
 
There have been thousands of 

602
00:34:29,679 --> 00:34:32,920
designs of slightly different 
 
airplanes or slightly different 

603
00:34:32,920 --> 00:34:35,280
cars and so on. 
 
So there is value in SimAI. 

604
00:34:38,120 --> 00:34:40,880
But then you ask the question, 

right, So where is the future? 

605
00:34:40,880 --> 00:34:49,199
The future is foundational 
 
models for AI where the customer

606
00:34:49,199 --> 00:34:54,600
will not have to train any set 

of things. 

607
00:34:54,600 --> 00:34:56,800
There is no need for a SimAI 
 
platform. 

608
00:34:57,840 --> 00:35:03,160
We, Ansys, will take the world 

of physics, of fluids around us 

609
00:35:03,720 --> 00:35:06,640
and we'll train it and that is 

what we will. 

610
00:35:07,000 --> 00:35:13,640
So we will train the AI, just 
 
like ChatGPT has trained all the

611
00:35:13,640 --> 00:35:15,320
words in the English language, 

right? 

612
00:35:15,640 --> 00:35:18,760
And has learned how to speak, 
 
how to write poetry. 

613
00:35:20,320 --> 00:35:24,920
The grand vision of AI with 
 
foundational models for physics 

614
00:35:24,920 --> 00:35:27,760
is to do that. 
 
It is an incredibly hard 

615
00:35:27,760 --> 00:35:30,480
problem, but that's what we are 
 working on. 

616
00:35:31,720 --> 00:35:34,760
But you, that's interesting 
 
because I've often had this 

617
00:35:34,760 --> 00:35:40,360
debate on the commercial or the 
 economics of that. 

618
00:35:41,120 --> 00:35:48,103
As in, if you are a car company,

 you probably have your own 

619
00:35:48,103 --> 00:35:50,352
cars. 
Like you said, it's quite 
 

620
00:35:50,360 --> 00:35:54,342
incremental and you'll train, 
you could train using your own 


621
00:35:54,350 --> 00:35:58,780
data that is proprietary to you.
And you, you would have assumed 

622
00:35:58,780 --> 00:36:01,942
 that the model you would train 
would be as accurate as possible

623
00:36:01,942 --> 00:36:05,345

 because it's your cars and 
your, your, your iterations. 
 

624
00:36:05,353 --> 00:36:09,035
Same if you're an aircraft 
designer. 
 

625
00:36:09,043 --> 00:36:15,395
I guess what you're alluding to 
is if your company or another 
 

626
00:36:15,403 --> 00:36:19,940
company could run their own 
simulations of all of these 
 

627
00:36:19,948 --> 00:36:24,802
different things and then train 
a massive model, will that model

628
00:36:24,802 --> 00:36:29,470

 be more accurate than the 
model that the car company has 

629
00:36:29,470 --> 00:36:32,686
trained 
 themselves? 
And. 
 

630
00:36:32,694 --> 00:36:35,303
Yes, yes, and and and here's 
why. 
 

631
00:36:35,311 --> 00:36:38,269
I'll go back to the Google 
example. 
 

632
00:36:38,277 --> 00:36:43,815
Like Google search is so good 
because it's a free tool, right?

633
00:36:43,815 --> 00:36:45,880

 
You and I type things on Google 

634
00:36:46,720 --> 00:36:49,520
and based on it, they are 
 
creating this massive database, 

635
00:36:49,520 --> 00:36:51,160
right? 
 
This PageRank algorithms of 

636
00:36:51,160 --> 00:36:53,920
this, tied to that and so on. 
 
And based on I'm clicking this, 

637
00:36:53,920 --> 00:36:58,080
right, gives you the list 
 of 
100 things, and you click this. 

638
00:36:58,080 --> 00:37:00,040
 
And the more you click, that is 

639
00:37:00,040 --> 00:37:04,360
a more important thing, right? 

And so if Google were to be only

640
00:37:04,360 --> 00:37:09,440
limited to the searches that 
 
Prith and Neil only did, that's 

641
00:37:09,440 --> 00:37:13,080
the only data they looked at, 
 
the search would not be as good.

642
00:37:14,280 --> 00:37:17,760
The reason Google is good is 
 
because they're looking at the 

643
00:37:17,760 --> 00:37:20,880
10 billion people on the planet 
 banging on our, on our 

644
00:37:20,880 --> 00:37:25,040
keyboards, right, for free. 
 
They, they, we think it is a 

645
00:37:25,040 --> 00:37:27,960
free thing. 
 
They're not paying us to give 

646
00:37:27,960 --> 00:37:31,880
us, give them the data, right? 

They're using all our 

647
00:37:31,880 --> 00:37:34,360
information to make the search 

be better. 

648
00:37:36,200 --> 00:37:42,520
So in exchange for us getting a 
 free tool, we are giving Google

649
00:37:42,520 --> 00:37:47,800
back the knowledge in our head 

that after I type who is Neil 

650
00:37:47,880 --> 00:37:51,080
Ashton from AWS, right? 
 
That somebody in the world is 

651
00:37:51,080 --> 00:37:54,000
actually interested in the 
 
question of who is Neil Ashton 

652
00:37:54,000 --> 00:37:56,560
from AWS? 
 
And the other question is who is

653
00:37:56,560 --> 00:37:58,198
Prith Banerjee from Ansys, 
right? 
 

654
00:37:58,206 --> 00:38:02,054
That is knowledge that is being 
captured by Google. 
 

655
00:38:02,062 --> 00:38:08,668
So now my knowledge is right. 
Suppose I went to train it only 

656
00:38:08,668 --> 00:38:14,005
 on Airbus designs, right? 
Think of it as the Google 
 

657
00:38:14,013 --> 00:38:17,965
search for only people within 
Airbus typing their searches 
 

658
00:38:17,973 --> 00:38:23,040
versus letting the searches go 
to all engineering companies, 
 

659
00:38:23,048 --> 00:38:26,666
right? 
To Airbus and Boeing and Pratt &

660
00:38:26,666 --> 00:38:31,680

 Whitney and GE, it will be 
clearly richer. 
 

661
00:38:31,688 --> 00:38:36,050
Yeah. 
That's the value now to make it 

662
00:38:36,050 --> 00:38:38,780
 happen, right? 
So, so I'll, I'll anticipate the

663
00:38:38,780 --> 00:38:41,800

 question, right. 
So where do you get the data 
 

664
00:38:41,808 --> 00:38:44,108
from, right? 
So this is something I'm 
 

665
00:38:44,116 --> 00:38:47,959
actually thinking of, right? 
So to build these foundational 


666
00:38:47,967 --> 00:38:53,320
models, I will have to get all 
the CAD files from Airbus and 
 

667
00:38:53,328 --> 00:38:58,510
all the CAD files from Boeing. 
Will Boeing and Airbus be 
 

668
00:38:58,518 --> 00:39:03,854
willing to give it to us? 
And this is the whole thing 
 

669
00:39:03,862 --> 00:39:05,464
about Google, right? 
I just give it. 
 

670
00:39:05,472 --> 00:39:08,840
The reason I give the Google 
example is they created a 
 

671
00:39:08,848 --> 00:39:13,525
business model which was free, 
perceived free, but in exchange 

672
00:39:13,525 --> 00:39:16,608
 for free they are sucking in 
their stuff, right? 
 

673
00:39:16,616 --> 00:39:23,600
So can I create a model which is
an opt-in model where all Ansys 

674
00:39:23,600 --> 00:39:29,460
 customers would opt to give 
their data in an anonymised way 

675
00:39:29,460 --> 00:39:34,015
 to Ansys to collect the data, 
run all those things for example

676
00:39:34,015 --> 00:39:39,685

 on the AWS cloud, right? 
And the cloud is a great way to 

677
00:39:39,685 --> 00:39:44,110
 train all these models because 
if it is on-prem, you actually 


678
00:39:44,118 --> 00:39:47,514
cannot have access to it. 
But if you are going on the 
 

679
00:39:47,522 --> 00:39:50,675
cloud, if every customer, if all
CAD designs are done on the 
 

680
00:39:50,683 --> 00:39:52,175
cloud, right, it is actually 
possible. 
 

681
00:39:52,183 --> 00:39:57,415
All you need is for Airbus and 
Boeing and Ford and GM to say 
 

682
00:39:57,423 --> 00:40:03,890
you can train the model. 
Just don't attribute it to Ford.

683
00:40:03,890 --> 00:40:05,200

 
Yeah, that's. 

684
00:40:06,360 --> 00:40:08,840
So now we are getting into 
 
policy-type things. 

685
00:40:09,040 --> 00:40:11,840
So it is possible. 
 
And so if if that were to 

686
00:40:11,840 --> 00:40:14,440
happen, it would be more 
 
accurate than the Airbus 

687
00:40:14,440 --> 00:40:17,560
specific result that is the long

 answer to that question. 

688
00:40:17,560 --> 00:40:19,960
Yeah, no, no. 
 
And I think this is actually 

689
00:40:19,960 --> 00:40:23,590
something that Max Welling, when

 I spoke with him, he brought 

690
00:40:23,590 --> 00:40:29,640
up which was the incentivizing 

people to share data, you know, 

691
00:40:29,840 --> 00:40:34,280
exactly, you know, that having 

some mechanism where either they

692
00:40:34,280 --> 00:40:37,760
get paid for it or they get 
 
something in return, something 

693
00:40:37,760 --> 00:40:40,680
that will allow them to overcome

 the barrier. 

694
00:40:41,000 --> 00:40:44,560
The, the sort of traditional 
 
position of this is our data. 

695
00:40:44,560 --> 00:40:46,938
I'm not going to, you know, let 
 anybody else use it to the 

696
00:40:46,938 --> 00:40:49,160
point where they see a benefit 
from 
 doing it. 

697
00:40:49,440 --> 00:40:54,000
I guess the technology piece is 
 the making it anonymous. 

698
00:40:54,080 --> 00:40:56,400
You know, that's probably the 
 
challenging bit is to sort of 

699
00:40:56,400 --> 00:40:59,360
figure out how to do it. 
 
But I guess this is not limited 

700
00:40:59,360 --> 00:41:01,640
to simulation. 
 
This is a broader question, 

701
00:41:01,640 --> 00:41:04,640
isn't it, to and? 
 
And for example, people are now 

702
00:41:05,160 --> 00:41:10,770
suing DALL-E right for hey, you,

 you are generating an image 

703
00:41:10,770 --> 00:41:14,440
based on my I'm an artist. 
 
I'm I'm whatever, right? 

704
00:41:15,040 --> 00:41:16,920
I'm Prith Banerjee. 
 
I've written a beautiful 

705
00:41:17,120 --> 00:41:19,640
picture, right? 
 
And you took that picture into 

706
00:41:19,640 --> 00:41:22,464
DALL-E and now you generate a 
new 
 picture, right, based on 

707
00:41:22,464 --> 00:41:26,040
its hidden knowledge. 
 
That's not fair if Prith 

708
00:41:26,040 --> 00:41:31,320
Banerjee were to say I will 
 
give 10 of my pictures to DALL-E

709
00:41:32,160 --> 00:41:37,160
or 10 of my poems to OpenAI 
 
right and I get one cent for 

710
00:41:37,160 --> 00:41:40,640
everything that I give every 
 
token I give to contribute to 

711
00:41:40,640 --> 00:41:44,680
this thing, I get one cent. 
 
Hey, I am I incentivised in that

712
00:41:44,680 --> 00:41:47,880
case, I will not sue DALL-E. 
 
So my thing is, I think the 

713
00:41:47,880 --> 00:41:52,720
whole world of GenAI, all the 
 
governance mechanism, so on, is 

714
00:41:52,800 --> 00:41:56,080
that people getting sued because

 they feel like their 

715
00:41:56,080 --> 00:41:58,680
intellectual property is not 
 
getting recognized, right? 

716
00:41:59,480 --> 00:42:02,680
Why do people have patents? 
 
Well, they have patents so that 

717
00:42:02,680 --> 00:42:04,880
they can have royalties based on

 the patents, right? 

718
00:42:05,320 --> 00:42:08,120
That royalties based on the 
 
patent, which is this big thing 

719
00:42:08,720 --> 00:42:11,640
in the world of AI. 
 
You have to figure it out how to

720
00:42:11,640 --> 00:42:16,300
take that big thing into small, 
 small chunks and to figure a 

721
00:42:16,300 --> 00:42:18,960
royalty of 1 cent per pixel, 
 
right? 

722
00:42:19,680 --> 00:42:21,680
Literally. 
 
I mean, the world will actually 

723
00:42:21,680 --> 00:42:25,040
go in that area. 
 
I I think the world will really 

724
00:42:25,040 --> 00:42:29,360
figure out it's governance and 

fairness so that everybody wins.

725
00:42:30,000 --> 00:42:32,920
Yeah. 
 
No, no, I, I think that's a 

726
00:42:32,960 --> 00:42:36,960
really good analogy. 
 
I, I, I and I agree with you. 

727
00:42:36,960 --> 00:42:40,200
I think if the data maybe this 

leads to the other one because 

728
00:42:40,200 --> 00:42:46,960
often the question is around 
 
data-driven versus physics- 

729
00:42:46,960 --> 00:42:51,280
driven with the logic being that

 you know, we operate in a 

730
00:42:51,280 --> 00:42:54,600
scientific world, we should 
 
include physics in the models. 

731
00:42:55,560 --> 00:42:58,080
But often the argument is we 
 
need to include physics because 

732
00:42:58,080 --> 00:43:02,680
we don't have enough data. 
 
And I have followed the progress

733
00:43:02,680 --> 00:43:04,840
of you know physics-informed 
 
etcetera. 

734
00:43:04,840 --> 00:43:10,200
But what struck me is to, to my 
 knowledge anyway, most of the 

735
00:43:10,280 --> 00:43:14,400
successful examples in the 
 
public domain have been with 

736
00:43:14,400 --> 00:43:20,680
data-driven approaches typically

 and not so much from the 

737
00:43:20,760 --> 00:43:25,320
theoretically better, but often 
 practically not as convenient. 

738
00:43:25,600 --> 00:43:28,720
So do you think that is just 
 
because it's harder and it will 

739
00:43:28,720 --> 00:43:32,720
take more time to sort of 
 
develop the more physics- 

740
00:43:32,720 --> 00:43:36,140
informed, physics-inspired, you 
 know, how much do you think 

741
00:43:36,140 --> 00:43:42,600
that is a needed science step to

 really overcome the data 

742
00:43:42,600 --> 00:43:45,920
challenge and the generalization

 challenge when it comes to AI 

743
00:43:45,920 --> 00:43:49,480
for, you know, computer-aided 
 
engineering? 

744
00:43:50,800 --> 00:43:53,120
That is a great question 
 
actually. 

745
00:43:53,120 --> 00:43:55,920
And the answer is there is not 

enough research that has 

746
00:43:55,920 --> 00:43:57,920
happened. 
 
There's more research that has 

747
00:43:57,920 --> 00:43:59,080
happened in the data world, 
 
right? 

748
00:44:00,000 --> 00:44:02,760
Purely data-driven method is 
 
more general. 

749
00:44:02,920 --> 00:44:05,560
That's the advantage, right? 
 
And and you don't need anything,

750
00:44:05,560 --> 00:44:07,960
just do do the data and you do 

your stuff, right. 

751
00:44:08,640 --> 00:44:12,360
But it requires enormous amount 
 of data, right to do the right 

752
00:44:12,360 --> 00:44:16,032
level of accuracy for the models

 where physics-informed gets 

753
00:44:16,032 --> 00:44:19,990
you is so so I mean, I just I 
know 
 you know this, but to 

754
00:44:19,990 --> 00:44:22,304
your readers, I will give you a 
very 
 simple explanation, 

755
00:44:22,304 --> 00:44:24,838
right? 
So suppose you are trying to 
 

756
00:44:24,846 --> 00:44:28,636
look at fluids data, right? 
And you are trying to put it 
 

757
00:44:28,644 --> 00:44:30,469
into an AI model for fluids, 
right? 
 

758
00:44:30,477 --> 00:44:33,915
You will take the the fluids 
data here is this velocity, 
 

759
00:44:33,923 --> 00:44:35,685
pressure, etcetera, temperature 
and so on. 
 

760
00:44:35,693 --> 00:44:36,890
And this is the distribution, 
right? 
 

761
00:44:36,898 --> 00:44:40,080
And you think the whole thing is
is random. 
 

762
00:44:40,088 --> 00:44:44,660
It is not because the fluids 
physics says there is Navier 
 

763
00:44:44,668 --> 00:44:47,410
Stokes equation, there is energy
conservation, all the stuff that

764
00:44:47,410 --> 00:44:49,680

 you know from a physics point 
of view. 
 

765
00:44:49,688 --> 00:44:53,840
So the data will not be 
completely uncorrelated. 
 

766
00:44:53,848 --> 00:44:59,480
The data is actually going to be
this thing with this turbulence 

767
00:44:59,480 --> 00:45:02,840
 will be constrained to only 
this set of things, right? 
 

768
00:45:02,848 --> 00:45:05,960
So if the, if there are three 
things here you are, you are 
 

769
00:45:05,968 --> 00:45:08,894
measuring, right? 
If it is, if you know these two 

770
00:45:08,894 --> 00:45:10,647
 points, you can deduce the 
third point. 
 

771
00:45:10,655 --> 00:45:13,215
It is actually not an 
independent variable, right? 
 

772
00:45:13,223 --> 00:45:16,684
That's the, the, the, the thing 
about statistics, right? 
 

773
00:45:16,692 --> 00:45:20,160
You think not all the things are
going to be independent. 
 

774
00:45:20,168 --> 00:45:23,109
So the pure data-driven approach
assumes everything is 
 

775
00:45:23,117 --> 00:45:26,950
independent and it's not. 
So if you can insert the 
 

776
00:45:26,958 --> 00:45:29,148
knowledge of the physics, you 
can constrain. 
 

777
00:45:29,156 --> 00:45:33,080
You say you don't have to search
for millions of data points, you

778
00:45:33,080 --> 00:45:35,927

 can do it with only 1000 data 
points. 
 

779
00:45:35,935 --> 00:45:37,943
That's the power of physics- 
informed, right? 
 

780
00:45:37,951 --> 00:45:41,304
And the work was, as you know, 
done by I mean, Karniadakis— 
 

781
00:45:41,312 --> 00:45:44,579
George Karniadakis— at Brown 
University, and we did a lot of 

782
00:45:44,579 --> 00:45:48,170
 work at Ansys we've done a lot 
of work at NVIDIA on these 
 

783
00:45:48,178 --> 00:45:50,536
things. 
The trouble is when we started 


784
00:45:50,544 --> 00:45:53,890
doing the physics-informed to 
incorporate the physics, the 
 

785
00:45:53,898 --> 00:45:58,589
computation needed in this, we 
have not quite figured that out,

786
00:45:58,589 --> 00:46:00,476

 right? 
And then then people went 
 

787
00:46:00,484 --> 00:46:03,985
towards graph neural networks 
and then NVIDIA did the work on 

788
00:46:03,985 --> 00:46:07,056
 FNO. 
So I saw this this lot of 
 

789
00:46:07,064 --> 00:46:10,392
research being done and 
somewhere in there in the Holy 


790
00:46:10,400 --> 00:46:14,780
Grail is foundational models. 
And once that thing is invented 

791
00:46:14,780 --> 00:46:19,210
 is like the Einstein theory of 
relativity, the the universal 
 

792
00:46:19,218 --> 00:46:23,860
thing, something like this will 
happen where we'll merge the 
 

793
00:46:23,868 --> 00:46:26,609
areas of numerical methods and 
AI. 
 

794
00:46:26,617 --> 00:46:32,880
And that I have told my board is
when the whole market for Ansys 

795
00:46:32,880 --> 00:46:36,140
 will completely collapse 
because we have the last 50 

796
00:46:36,140 --> 00:46:39,105
years we 
 have worked on, on 
the fact that it's all numerical

797
00:46:39,105 --> 00:46:42,618
methods and 
 so on, right? 
Numerical methods will no longer

798
00:46:42,618 --> 00:46:45,942

 be needed. 
It will all be done with AI, 
 

799
00:46:45,950 --> 00:46:50,064
with the accuracy and the speed 
much, much better than than 
 

800
00:46:50,072 --> 00:46:52,080
numerical methods. 
But we're not there yet. 
 

801
00:46:52,088 --> 00:46:53,538
That's where the research is 
needed. 
 

802
00:46:53,546 --> 00:46:59,080
And that actually brings me on 
to the point then of how, how 
 

803
00:46:59,088 --> 00:47:01,852
can we enable that research to 
happen. 
 

804
00:47:01,860 --> 00:47:06,580
As you said yourself, a very 
large enterprise will struggle 


805
00:47:06,588 --> 00:47:10,800
to dedicate resources to 
fundamental problems because of 

806
00:47:10,800 --> 00:47:14,136
 the pressures of headcount, of 
incremental product improvement.

807
00:47:14,136 --> 00:47:17,720

 
A start up can do that, but 

808
00:47:17,720 --> 00:47:20,640
they're not, they have pressure 
 from their VCs to actually 

809
00:47:20,640 --> 00:47:23,628
deliver something within a, you 
 know, relatively small amount 

810
00:47:23,628 --> 00:47:26,960
of time usually. 
 
So it falls down to academia. 

811
00:47:27,880 --> 00:47:34,160
But if, let's say foundational 

models is, as you rightly say, 

812
00:47:34,480 --> 00:47:37,400
could be a, you know, Eureka 
 
moment, you know, a massive 

813
00:47:37,400 --> 00:47:41,240
moment for the field. 
 
The bit that I've noticed is 

814
00:47:41,240 --> 00:47:45,640
data, you know, you could 
 
incentivize people to give you 

815
00:47:45,640 --> 00:47:48,960
data through, you know, and CIS 
 and, and mechanisms. 

816
00:47:49,480 --> 00:47:53,800
But I wonder, therefore, what's 
 your opinion of the open source

817
00:47:53,800 --> 00:47:56,680
versus closed source? 
 
You know, how much should we be 

818
00:47:56,680 --> 00:47:59,160
trying to create some open 
 
source data that's to help the 

819
00:47:59,160 --> 00:48:02,720
academia, but not do too much 
 
that you give all your IP away 

820
00:48:02,720 --> 00:48:04,760
and you know, you, you, you lose

 an advantage. 

821
00:48:04,760 --> 00:48:08,600
So where does that balance? 
 
That is a great point. 

822
00:48:08,600 --> 00:48:14,760
In fact, let me tell you, the 
 
reason that the AI world has 

823
00:48:17,040 --> 00:48:20,840
worked so fast is because of 
 
open source, right? 

824
00:48:21,680 --> 00:48:24,956
You have things like TensorFlow 
 and PyTorch, these are 

825
00:48:24,956 --> 00:48:29,600
absolutely open source ways of 
doing neural 
 networks, right? 

826
00:48:29,600 --> 00:48:32,840
I mean, you, they could have 
 
like Google and so on could have

827
00:48:32,840 --> 00:48:35,360
kept all of those and Facebook 

could have kept it completely 

828
00:48:35,360 --> 00:48:38,360
closed, right? 
 
And then the world wouldn't have

829
00:48:38,520 --> 00:48:39,840
done all this kind of stuff, 
 
right? 

830
00:48:40,960 --> 00:48:44,880
NVIDIA opened up CUDA, right? 
 
So CUDA became sort of not, and 

831
00:48:44,880 --> 00:48:47,240
the code is not open. 
 
So, but they have open 

832
00:48:47,280 --> 00:48:49,720
framework, right? 
 
So the combination of open 

833
00:48:49,720 --> 00:48:54,760
source things like CUDA, open 
 
source things like PyTorch and 

834
00:48:54,760 --> 00:49:01,160
TensorFlow, etcetera, and open 

source models for the data like 

835
00:49:01,160 --> 00:49:03,960
ImageNet and all those things 
 
that are out there, right? 

836
00:49:04,520 --> 00:49:09,080
It has accelerated the pace of 

innovation in the world of AI, 

837
00:49:09,680 --> 00:49:13,520
unlike other fields like in the 
 world of numerical methods, we 

838
00:49:13,520 --> 00:49:15,560
at Ansys know something, 
Dassault 
 knows something. 

839
00:49:15,560 --> 00:49:17,633
Siemens knows something and we 
don't sort of 
 share stuff, 

840
00:49:17,633 --> 00:49:20,186
right. 
So there's a paper they'll come 

841
00:49:20,186 --> 00:49:24,110
 from CMU or Stanford, some 
wonderful people and we, we say,

842
00:49:24,110 --> 00:49:27,640

 ah, but they are working on, 
on trivial toy problems. 
 

843
00:49:27,648 --> 00:49:31,660
They cannot work on Ansys, but 
Ansys will not give those tough 

844
00:49:31,660 --> 00:49:35,135
 problems that did not happen in
the AI world, that did not 
 

845
00:49:35,143 --> 00:49:38,788
happen in the MapReduce world, 
in the MapReduce world, right? 


846
00:49:38,796 --> 00:49:44,240
The MapReduce thing was actually
openly given away, 
 right? 

847
00:49:44,240 --> 00:49:48,200
Open source by both Google and 

and Yahoo. 

848
00:49:48,640 --> 00:49:53,320
Now why did they do that? 
 
That MapReduce framework is a 

849
00:49:53,320 --> 00:49:57,880
framework that Google needs to 

improve their searches, right? 

850
00:49:57,880 --> 00:50:01,914
So it was a brilliant business 

move for them to open-source 

851
00:50:01,914 --> 00:50:05,304
Map- Reduce to the world, right 
where 
 the smartest graduate 

852
00:50:05,304 --> 00:50:10,074
students at MIT and Oxford and 
and CMU, 
 they all work to 

853
00:50:10,074 --> 00:50:12,220
improve Map- Reduce and it's 
open source. 
 

854
00:50:12,228 --> 00:50:15,828
So all the innovations that 
coming from the the open 
 

855
00:50:15,836 --> 00:50:20,040
source world academic world, 
Google could put in and make the

856
00:50:20,040 --> 00:50:23,675

 search even better, right? 
They did not say here is a 
 

857
00:50:23,683 --> 00:50:27,200
search algorithm that we open 
source that they took a core 
 

858
00:50:27,208 --> 00:50:31,508
part of their search algorithm 
which they are making money off,

859
00:50:31,508 --> 00:50:35,667

 right, with ads. 
So I thought that was an 
 

860
00:50:35,675 --> 00:50:38,360
absolute brilliant strategy. 
Linux is another brilliant 
 

861
00:50:38,368 --> 00:50:41,080
strategy, right, for for 
advancing operating systems, 
 

862
00:50:41,088 --> 00:50:44,090
right? 
Which is so we have to actually 

863
00:50:44,090 --> 00:50:48,234
 learn. 
So the in our world of CAE 
 

864
00:50:48,242 --> 00:50:52,448
simulation, right, there is 
obviously one code called 
 

865
00:50:52,456 --> 00:50:54,820
OpenFOAM. 
Since you know, fluids, you know

866
00:50:54,820 --> 00:50:58,790

 OpenFOAM, right? 
So it's open source, actually 
 

867
00:50:58,798 --> 00:51:03,020
OpenFOAM plus OpenAI. 
I mean sort of is sort of where 

868
00:51:03,020 --> 00:51:06,874
 I think things will happen. 
But for that you also need that 

869
00:51:06,874 --> 00:51:11,026
 data for the CAD models, right?
And so just like ImageNet has 
 

870
00:51:11,034 --> 00:51:16,542
created this thing for 2D and 3D
images, we need in this area 
 

871
00:51:16,550 --> 00:51:21,662
some work on 3D geometries of 
all kinds of things, on gears 
 

872
00:51:21,670 --> 00:51:24,280
and this and and propellers and 
airplanes and so on. 
 

873
00:51:24,288 --> 00:51:27,250
If you can do that, I think that
will advance the 
 

874
00:51:27,258 --> 00:51:29,315
state-of-the-art. 
Yeah. 
 

875
00:51:29,323 --> 00:51:34,035
And we, you know, we, we 
ourselves published a couple of 

876
00:51:34,035 --> 00:51:35,846
 datasets: DrivAerML and 
AhmedML. 
 

877
00:51:35,854 --> 00:51:40,882
That I I am aware so. 
The which have helped a little 


878
00:51:40,890 --> 00:51:44,282
bit, but I. 
Not at the level of ImageNet, 
 

879
00:51:44,290 --> 00:51:47,080
not at the level of of. 
Exactly. 
 

880
00:51:47,088 --> 00:51:50,243
Exactly. 
But the the other bit that I 
 

881
00:51:50,251 --> 00:51:55,120
always and maybe I just need to 
get my brain around this, which 

882
00:51:55,120 --> 00:52:00,175
 is if you're training like a 
large language model, the text 


883
00:52:00,183 --> 00:52:05,100
and the data that you scrape up 
the Internet is in some ways 
 

884
00:52:05,108 --> 00:52:09,845
it's sort of the ground truth. 
Yes, someone has written it, but

885
00:52:09,845 --> 00:52:13,620

 it isn't a simulation, right? 
It is someone writing it. 
 

886
00:52:13,628 --> 00:52:18,042
If you transfer now to, let's 
say CFD, I could run simulations

887
00:52:18,042 --> 00:52:24,200

 of thousands of cars and 
planes, but the model will only,

888
00:52:24,200 --> 00:52:29,500
well, 
 this is my question. 
Using, let's say, just a 
 

889
00:52:29,508 --> 00:52:33,620
standard approach, it's only 
going to learn the equivalent 
 

890
00:52:33,628 --> 00:52:38,085
simulation settings if it's a 
RANS approach or an LES 
 

891
00:52:38,093 --> 00:52:41,232
approach, or a mesh that is 
coarse or fine. 
 

892
00:52:41,240 --> 00:52:46,109
If I do all my simulations with 
a RANS, the model's going to 
 

893
00:52:46,117 --> 00:52:52,477
learn. 
So does that mean you have a 
 

894
00:52:52,485 --> 00:52:56,472
foundational model of this 
simulation approach and then you

895
00:52:56,472 --> 00:52:57,840

 have another foundational 
model? 

896
00:52:57,840 --> 00:53:01,120
Or is there some way? 
 
Or or you have to curate your 

897
00:53:01,120 --> 00:53:06,080
data so that you take 10% of the

 data from RANS, 10% data from 

898
00:53:06,120 --> 00:53:09,680
LES, 10% data from DNS. 
 
That would be the real 

899
00:53:09,680 --> 00:53:12,240
foundational model. 
 
So getting back to your issue, 

900
00:53:12,240 --> 00:53:14,680
right? 
 
And so the reason, so I'm glad 

901
00:53:14,680 --> 00:53:18,360
you asked this question because 
 with large language models for 

902
00:53:18,360 --> 00:53:20,840
words as tokens, you are 
 
absolutely right. 

903
00:53:20,840 --> 00:53:25,200
They have scraped all the words 
 from all the books that people 

904
00:53:25,200 --> 00:53:28,000
have written, right? 
 
Of course, they are not giving 

905
00:53:28,000 --> 00:53:29,960
the royalty back to the people 

who have written those things, 

906
00:53:29,960 --> 00:53:35,120
right, Pop chap, right. 
 
They have not generated those. 

907
00:53:35,360 --> 00:53:40,440
Suppose a next version of Chat- 
 GPT is take ChatGPT to generate

908
00:53:40,720 --> 00:53:43,360
all those texts, right? 
 
And then you feed it, that would

909
00:53:43,360 --> 00:53:44,720
be what your problem would be, 

right? 

910
00:53:45,360 --> 00:53:49,320
So and there and, and DALL-E has

 taken the same approach of 

911
00:53:49,320 --> 00:53:52,920
taking all the images on the 
 
Internet and use those images to

912
00:53:52,920 --> 00:53:56,120
program DALL-E, right? 
 
And Sora has done the same thing

913
00:53:56,120 --> 00:54:00,680
for for videos, right? 
 
So in our world to do the 

914
00:54:00,680 --> 00:54:03,440
foundational models, it has to 

be the 3D fields. 

915
00:54:03,680 --> 00:54:07,920
Now you can actually go and 
 
measure 3D fields, right? 

916
00:54:08,200 --> 00:54:13,080
You you take a car right Google 
 Waymo those cars have videos 

917
00:54:13,080 --> 00:54:15,800
right. 
 
So imagine you you sensorise 

918
00:54:15,800 --> 00:54:21,680
your car to measure the fluid 
 
flow at every small microsecond 

919
00:54:21,680 --> 00:54:24,400
or micro whatever millimeter of 
 your car right. 

920
00:54:24,760 --> 00:54:28,400
That is an actual measurement. 

I mean this is the air how the 

921
00:54:28,400 --> 00:54:31,520
airflow actually happened on the

 on the car right that you have

922
00:54:31,520 --> 00:54:34,360
to take thousands of cars 
 
millions of cars and so on. 

923
00:54:34,360 --> 00:54:37,840
It's just ridiculous right. 
 
So what I am saying what I have 

924
00:54:37,840 --> 00:54:41,920
told my board is we Ansys will 

create the synthetic data 

925
00:54:43,080 --> 00:54:46,480
through simulation right of all 
 the fluids model. 

926
00:54:46,480 --> 00:54:50,408
But you are absolutely right, we

 Ansys Fluent is RANS 

927
00:54:50,408 --> 00:54:52,515
simulation. 
So it will not be generating the

928
00:54:52,515 --> 00:54:55,160

 LES. 
So we will also have to do the 


929
00:54:55,168 --> 00:54:59,376
LES and the DNS and for external
fluid and for for cars and 
 

930
00:54:59,384 --> 00:55:03,344
trucks and so on. 
So only if you do all of that 
 

931
00:55:03,352 --> 00:55:08,462
will be a true foundational 
model, which is why if ChatGPT 


932
00:55:08,470 --> 00:55:14,632
took six months and 1.8 trillion
parameters, in our world, it is 

933
00:55:14,632 --> 00:55:16,555
 probably a billion trillion 
parameter. 
 

934
00:55:16,563 --> 00:55:20,125
I don't even know what the size 
of the model is, but it is 
 

935
00:55:20,133 --> 00:55:22,496
possible. 
I absolutely it is possible and 

936
00:55:22,496 --> 00:55:25,694
 we'll eventually get there. 
Yeah, it is. 
 

937
00:55:25,702 --> 00:55:30,240
It is so fascinating, isn't it? 
Because that it would be a 
 

938
00:55:30,248 --> 00:55:32,600
transformational change. 
Like you say, you're right to 
 

939
00:55:32,608 --> 00:55:36,035
tell your board that it's the 
honest truth that the tradition 

940
00:55:36,035 --> 00:55:40,340
 of, you know, running your own 
simulations, if the model was 
 

941
00:55:40,348 --> 00:55:45,179
accurate enough and you know, 
there's a big if I guess on 
 

942
00:55:45,187 --> 00:55:49,302
that, it would certainly become 
a very widely, it would disrupt 

943
00:55:49,302 --> 00:55:52,820
 the market, that's for sure in 
a, in a big, big way. 
 

944
00:55:52,828 --> 00:55:56,130
Let me. 
Let me make this statement to 
 

945
00:55:56,138 --> 00:55:59,449
your readers. 
I know you know this, but the 
 

946
00:55:59,457 --> 00:56:04,815
whole world of physics, right? 
When Newton observed an apple, 


947
00:56:04,823 --> 00:56:09,114
he dropped it and it fell down 
and he derived something else, 


948
00:56:09,122 --> 00:56:12,224
right? 
Just by a bunch of observations 

949
00:56:12,224 --> 00:56:17,580
 in the real world. 
That data that he fed into his 


950
00:56:17,588 --> 00:56:23,656
engine right determined the law 
which is force equals mass times

951
00:56:23,656 --> 00:56:28,448

 acceleration which is a 
differential equation right. 
 

952
00:56:28,456 --> 00:56:34,994
He deduced the law of gravity by
observations. 
 

953
00:56:35,002 --> 00:56:42,580
It is therefore possible to 
observe the world around us and 

954
00:56:42,580 --> 00:56:46,800
 actually we have got work going
on with Google DeepMind right. 


955
00:56:46,808 --> 00:56:52,054
DeepMind is using the Ansys 
tools to observe the physics and

956
00:56:52,054 --> 00:56:56,676

 learn the physics. 
So that imagine you're trying to

957
00:56:56,676 --> 00:56:59,978

 balance a long pen on your 
head, right? 
 

958
00:56:59,986 --> 00:57:01,940
Yeah, I'm sure when you're a 
kid, you did that and you're 
 

959
00:57:01,948 --> 00:57:04,616
balancing, right. 
When you ball, the thing goes on

960
00:57:04,616 --> 00:57:07,160

 the other side. 
You you move your hand, right. 


961
00:57:07,168 --> 00:57:11,262
So they have actually taken that
as an example to use Ansys 
 

962
00:57:11,270 --> 00:57:14,830
Mechanical to model the world of
structures, right? 
 

963
00:57:14,838 --> 00:57:20,525
And just by say and thereby 
train the Google DeepMind 
 has 

964
00:57:20,525 --> 00:57:26,120
trained the robot arm to do the 
balancing of this by 
 learning 

965
00:57:26,120 --> 00:57:29,410
the physics. 
So it is possible and that is 
 

966
00:57:29,418 --> 00:57:33,960
how I think foundational models 
will work because Isaac Newton 


967
00:57:33,968 --> 00:57:38,584
generated the physics model of 
force equals mass times 
 

968
00:57:38,592 --> 00:57:40,795
acceleration by looking at the 
data. 
 

969
00:57:40,803 --> 00:57:45,565
By observing the data, AI is 
going to observe the physics 
 

970
00:57:45,573 --> 00:57:48,606
around us and train the physics 
models. 
 

971
00:57:48,614 --> 00:57:52,188
Every one of those equations can
be deduced. 
 

972
00:57:52,196 --> 00:57:55,920
Navier–Stokes equations can be 
reverse engineered by AI. 
 

973
00:57:55,928 --> 00:57:59,280
And that would be, it's 
interesting to bring up the 
 

974
00:57:59,288 --> 00:58:02,179
robotics angle to that, because 
I guess this is the, the, the 
 

975
00:58:02,187 --> 00:58:05,224
bigger picture, isn't it the 
sort of future of manufacturing,

976
00:58:05,224 --> 00:58:08,219

 the future of robotics? 
You know, CAE, you don't just 
 

977
00:58:08,227 --> 00:58:10,458
simulate for the sake of it, do 
you? 
 

978
00:58:10,466 --> 00:58:14,860
You simulate it to do something.
And so you're right, there is a 

979
00:58:14,860 --> 00:58:17,720
 much, you know, broader, 
broader picture around. 
 

980
00:58:17,728 --> 00:58:23,565
However, one thing a sort of 
counter example to that, I 
 

981
00:58:23,573 --> 00:58:26,920
guess, which goes back to your, 
well, not a counter example, but

982
00:58:26,920 --> 00:58:28,212

 another way of thinking about 
it. 
 

983
00:58:28,220 --> 00:58:33,168
You said right at the beginning,
accuracy and speed or cost, It's

984
00:58:33,168 --> 00:58:37,100

 true that all engineering 
companies are so, you know, 
 

985
00:58:37,108 --> 00:58:41,127
focused on that, aren't they? 
Accuracy, speed, cost. 
 

986
00:58:41,135 --> 00:58:45,280
So if your traditional 
simulation could be fast enough 

987
00:58:45,280 --> 00:58:48,772
 and cheap enough, you don't 
necessarily need AI, do you? 
 

988
00:58:48,780 --> 00:58:54,224
That could be a normal approach.
So I was just wondering the 
 

989
00:58:54,232 --> 00:58:58,988
quantum piece, everybody brings 
this up and I would love to get 

990
00:58:58,988 --> 00:59:03,042
 your perspective on how 
realistic is it and of what 
 

991
00:59:03,050 --> 00:59:06,450
parts of you think that CAE 
could speed up the quantum side 

992
00:59:06,450 --> 00:59:08,864
 of things? 
Could could quantum speed up the

993
00:59:08,864 --> 00:59:13,960

 CAE side of things? 
Or do you see it still as being 

994
00:59:13,960 --> 00:59:18,262
 too niche and yeah, not? 
No, no, it's a great question. 


995
00:59:18,270 --> 00:59:21,760
And actually in my CTO office, 
so the the first thing I did in 

996
00:59:21,760 --> 00:59:26,321
 my CTO office was to work on AI
and now I've got now that that's

997
00:59:26,321 --> 00:59:29,365

 AI thing is sort of not 
solved, but at least we have 

998
00:59:29,365 --> 00:59:30,954
some 
 products out in this 
area, right? 
 

999
00:59:30,962 --> 00:59:35,020
We have started working on 
quantum for exactly that reason.

1000
00:59:35,020 --> 00:59:37,680

 
And the the beauty of quantum is

1001
00:59:37,880 --> 00:59:40,760
it is a potential for 
 
exponential speedups, right? 

1002
00:59:41,560 --> 00:59:45,800
Because if you have N qubits, 
 
your speed, your, your, your 

1003
00:59:45,800 --> 00:59:48,400
runtime, your, your speed up is 
 2ᴺ, right? 

1004
00:59:48,840 --> 00:59:52,680
And so as you typically quantum 
 computing algorithms have been 

1005
00:59:52,680 --> 00:59:56,840
used on problems that are sort 

of NP-complete to begin with, 

1006
00:59:56,840 --> 01:00:00,440
right, exponential problems like

 materials discovery, your 

1007
01:00:00,440 --> 01:00:02,560
optimisation and travelling 
 
salesman and so on. 

1008
01:00:03,200 --> 01:00:06,520
And you would think that in our 
 world, right our our problems 

1009
01:00:06,520 --> 01:00:10,880
are polynomial: N cubed. 
 
Except that N is huge, N is a 

1010
01:00:10,880 --> 01:00:13,080
million million cubed is a large

 number, right? 

1011
01:00:13,360 --> 01:00:19,200
So if you can throw a 2 to the 

power P kind of exponential 

1012
01:00:19,200 --> 01:00:23,600
capacity to solve a polynomial 

problem order N cubed where N is

1013
01:00:23,600 --> 01:00:26,640
very large, yes there are 
 
benefits. 

1014
01:00:26,920 --> 01:00:31,920
So we are looking at ways these 
 algorithms called the HHL 

1015
01:00:31,920 --> 01:00:36,440
algorithm that you may have 
 
heard of that can take a because

1016
01:00:36,440 --> 01:00:40,640
in all our things like we 
 
ultimately we will take our PDEs

1017
01:00:40,640 --> 01:00:46,560
and and make it into some Ax 
 =
b, some matrix vector. 

1018
01:00:46,640 --> 01:00:48,960
So you are trying to solve some 
 linear system equations. 

1019
01:00:49,520 --> 01:00:54,120
Turns out that quantum computers

 can solve linear systems of 

1020
01:00:54,120 --> 01:00:56,240
equations like with exponential 
 speed up. 

1021
01:00:56,240 --> 01:00:59,680
So we are looking into those 
 
kind of methods at Ansys. 

1022
01:01:00,280 --> 01:01:03,138
It is not going to be next year.

 

1023
01:01:03,146 --> 01:01:07,080
It'll not be two years, but 
definitely within 10 years we'll

1024
01:01:07,080 --> 01:01:10,064

 see quantum computing 
accelerating CAE simulation. 
 

1025
01:01:10,072 --> 01:01:15,316
And that'll be interesting. 
It's almost like you never quite

1026
01:01:15,316 --> 01:01:19,248
clear 
 what technology will be 
the one that transforms. 
 

1027
01:01:19,256 --> 01:01:22,865
You know, the I know it's not a 
great analogy, but you don't 
 

1028
01:01:22,873 --> 01:01:25,740
remember 3D glasses. 
The televisions came out and 
 

1029
01:01:25,748 --> 01:01:27,274
everyone thought that would be 
it. 
 

1030
01:01:27,282 --> 01:01:31,532
We'll all wear 3D glasses. 
And at least to my knowledge, it

1031
01:01:31,532 --> 01:01:35,344

 sort of died away because in 
reality, nobody wants to put 
 

1032
01:01:35,352 --> 01:01:37,862
those on. 
Now maybe there's a future one, 

1033
01:01:37,862 --> 01:01:39,807
 you know, with the Apple 
vision, etcetera. 
 

1034
01:01:39,815 --> 01:01:43,882
But it's amazing how resilient 
we have been to watching a 
 

1035
01:01:43,890 --> 01:01:47,551
normal TV, even with all the 
sort of technology changes. 
 

1036
01:01:47,559 --> 01:01:53,580
So it does often make me wonder,
is it ML, is it quantum? 
 

1037
01:01:53,588 --> 01:01:57,420
You know will one of them. 
And actually this quantum ML 
 

1038
01:01:57,428 --> 01:02:00,067
people are now working on 
quantum machine learning. 
 

1039
01:02:00,075 --> 01:02:03,550
So, so, so it is this 
combination of things that means

1040
01:02:03,550 --> 01:02:08,512

 to your broader question, what
I I would say is Ansys is now a 

1041
01:02:08,512 --> 01:02:11,880
 company of about 2 1/2 billion 
dollars, right? 
 

1042
01:02:11,888 --> 01:02:16,360
And our software originated from
CAE simulation written in 

1043
01:02:16,360 --> 01:02:20,215
Fortran 
 in 1970, right? 
But then as newer technologies 


1044
01:02:20,223 --> 01:02:23,318
like HPC came in, we took that 
Fortran code and said, OK, let's

1045
01:02:23,318 --> 01:02:26,248

 put it on a shared memory and 
with this directive, parallelise

1046
01:02:26,248 --> 01:02:28,699

 it. 
And then the GPUs came in, well,

1047
01:02:28,699 --> 01:02:31,420

 with this directive, make it 
run on CUDA and so on. 
 

1048
01:02:31,428 --> 01:02:35,050
So it has been always 
retrofitting a piece of thing. 


1049
01:02:35,058 --> 01:02:38,576
And somehow we have OK, now 
let's work with this. 
 

1050
01:02:38,584 --> 01:02:42,287
And now the AI/ML came in and 
let's do it with with 
 

1051
01:02:42,295 --> 01:02:43,682
TensorFlow. 
And now let this cloud came in. 

1052
01:02:43,682 --> 01:02:44,840
 
Let's try to make it on the 

1053
01:02:44,840 --> 01:02:47,520
cloud, right? 
 
What if, and this is sort of a 

1054
01:02:47,520 --> 01:02:50,000
thought experiment. 
 
I, I asked of my technology, I 

1055
01:02:50,000 --> 01:02:54,985
said, what if you knew that you 
 have access to quantum, to 

1056
01:02:54,985 --> 01:03:00,204
AI/ML, to cloud, to GPUs, all of
those 
 technologies and you to 

1057
01:03:00,204 --> 01:03:03,840
start writing Ansys Mechanical, 
How 
 would you write it? 

1058
01:03:04,200 --> 01:03:07,880
You would definitely and you 
 
have languages like Julia and 

1059
01:03:07,880 --> 01:03:11,440
Python And so on, right? 
 
Would you write it in Fortran 

1060
01:03:11,560 --> 01:03:13,280
with the linear algebra? 
Not at all 
 right. 

1061
01:03:14,880 --> 01:03:19,480
And this is sort of the 
 
advantage that startups have. 

1062
01:03:19,520 --> 01:03:23,680
A startup has no legacy. 
 
This is what how I try to 

1063
01:03:23,680 --> 01:03:26,240
motivate people. 
 
I said if you are a startup, you

1064
01:03:26,240 --> 01:03:31,117
have the latest widgets, right 

With with AR, VR, IoT—I don't 

1065
01:03:31,117 --> 01:03:33,520
know what it is right. 
 
Put all of that in the blender 

1066
01:03:33,520 --> 01:03:37,440
and out will come something that

 a large company like Ansys has

1067
01:03:37,440 --> 01:03:39,240
not had the luxury of doing 
 
right. 

1068
01:03:39,840 --> 01:03:41,680
That's the power of the Horizon 
 3. 

1069
01:03:42,840 --> 01:03:46,040
And that that's a great sort of 
 circle back to our original 

1070
01:03:46,040 --> 01:03:50,360
point that essentially I think 

what you're saying is you want 

1071
01:03:50,440 --> 01:03:53,880
the startups to take that 
 
challenge to say, I want to 

1072
01:03:53,880 --> 01:03:58,680
start from scratch. 
 
Prove to me, Ansys, that a new 

1073
01:03:58,680 --> 01:04:04,240
written code using the latest 
 
technology right now could be, 

1074
01:04:04,360 --> 01:04:07,840
you know, exponentially better 

than retrofitting another code. 

1075
01:04:08,280 --> 01:04:10,720
And that is what would make a 
 
startup attractive to you. 

1076
01:04:10,720 --> 01:04:12,160
That's what would change the 
 
market. 

1077
01:04:12,160 --> 01:04:14,600
That's what would do it. 
 
And it's only really a startup. 

1078
01:04:14,640 --> 01:04:17,000
And the business model is the 
 
following, right? 

1079
01:04:17,920 --> 01:04:21,800
Oftentimes I am asked, at 
 
Ansys, hey, you are doing this 

1080
01:04:21,800 --> 01:04:23,640
right with AI/ML, why are you 
 
doing this? 

1081
01:04:23,680 --> 01:04:25,200
Is you'll kill your cash cow, 
 
right? 

1082
01:04:25,440 --> 01:04:28,880
Because if with AI you know 
 
things run 100 times faster, 

1083
01:04:28,880 --> 01:04:32,360
it's not good for the Ansys 
 
Mechanical business or fluid 

1084
01:04:32,360 --> 01:04:36,640
business, right? 
 
But the response is if I don't 

1085
01:04:36,640 --> 01:04:40,360
do it myself, a startup would do

 it and destroy me anyway. 

1086
01:04:40,480 --> 01:04:42,400
So I might as well do it myself,

 right. 

1087
01:04:42,520 --> 01:04:46,120
So it is like the classic case 

of Kodak. 

1088
01:04:47,480 --> 01:04:51,040
They actually knew of digital 
 
printing. 

1089
01:04:52,120 --> 01:04:53,480
It's not like those guys are 
 
stupid. 

1090
01:04:53,480 --> 01:04:54,560
They actually know digital 
 
thing. 

1091
01:04:54,560 --> 01:04:58,160
But the cash cow from analog was

 so good that they did not want

1092
01:04:58,160 --> 01:05:01,960
to disrupt it, right? 
 
Motorola knew about digital 

1093
01:05:01,960 --> 01:05:06,200
phones, but they, they, they 
 
were afraid of it, right? 

1094
01:05:06,200 --> 01:05:09,120
So it's always the innovator's 

dilemma, right? 

1095
01:05:09,240 --> 01:05:10,720
I have a cash cow business, 
 
right? 

1096
01:05:11,040 --> 01:05:14,240
Should I do it? 
 
And so companies like Apple 

1097
01:05:15,040 --> 01:05:19,951
where Steve Jobs said the iPhone

 is going to disrupt iPod, but 

1098
01:05:19,951 --> 01:05:23,096
I would rather disrupt iPod 
myself 
 than be disrupted with 

1099
01:05:23,096 --> 01:05:26,040
somebody else. 
 
And whereas a startup has got 

1100
01:05:26,040 --> 01:05:28,000
nothing to lose, right? 
 
They're starting with zero 

1101
01:05:28,000 --> 01:05:30,920
revenue, right? 
 
So that's the advantage of a 

1102
01:05:30,920 --> 01:05:33,360
startup. 
 
And what I write in my book is 

1103
01:05:33,360 --> 01:05:38,320
therefore companies like Ansys, 
 like Kodak, like Motorola need 

1104
01:05:38,320 --> 01:05:42,400
to work with the digital 
 
printers, the digital phones, 

1105
01:05:42,400 --> 01:05:47,464
the digital, the quantum or 
AI-based 
 simulation and 

1106
01:05:47,464 --> 01:05:51,080
embrace them and bring them into
your thing, 
 right? 

1107
01:05:51,080 --> 01:05:55,080
That's the way a company like 
 
Ansys can actually stay on top 

1108
01:05:55,080 --> 01:05:56,647
of Horizon 3 Innovations, right?

 

1109
01:05:56,655 --> 01:05:59,570
Yeah. 
And I think that's probably 
 

1110
01:05:59,578 --> 01:06:03,815
where I think that, you know, 
Jeff Bezos with his original 
 

1111
01:06:03,823 --> 01:06:06,729
Amazon leadership principles, 
the one of customer obsession 
 

1112
01:06:06,737 --> 01:06:11,728
makes sense if you just focus on
what, what would a customer want

1113
01:06:11,728 --> 01:06:15,054

 that typically always works. 
And that always works. 
 

1114
01:06:15,062 --> 01:06:17,320
That always. 
And that has been my 
 defining 

1115
01:06:17,320 --> 01:06:18,755
philosophy. 
Yeah. 
 

1116
01:06:18,763 --> 01:06:23,951
Well, maybe to close out, I 
would love just to get some of 


1117
01:06:23,959 --> 01:06:26,154
your, you know, summarised 
advice. 
 

1118
01:06:26,162 --> 01:06:29,680
I guess there's people listening
to this who are maybe, you know,

1119
01:06:29,680 --> 01:06:32,461

 coming to the end of their 
PhDs, They're in industry. 
 

1120
01:06:32,469 --> 01:06:35,165
They have ideas. 
What would be your if you had to

1121
01:06:35,165 --> 01:06:38,048

 summarize? 
What have helped you to get to 


1122
01:06:38,056 --> 01:06:40,615
such an illustrious position 
where, where you are now? 
 

1123
01:06:40,623 --> 01:06:44,258
How What advice would you give 
to people at the end of this to 

1124
01:06:44,258 --> 01:06:46,463
 motivate them in their in their
careers? 
 

1125
01:06:46,471 --> 01:06:50,986
So, so the motivational thing is
exactly what I kind of covered 


1126
01:06:50,994 --> 01:06:54,761
in the last minute, right? 
That they have to just in fact, 

1127
01:06:54,761 --> 01:06:58,435
 in my book, I talk about the 
digital technology that we have 

1128
01:06:58,435 --> 01:07:00,420
 today, right? 
I talk about quantum, I talk 
 

1129
01:07:00,428 --> 01:07:03,068
about AI, talk about IoT, talk 
about platforms, all the stuff 


1130
01:07:03,076 --> 01:07:07,631
that is there, right? 
So here is a kid graduating with

1131
01:07:07,631 --> 01:07:11,564

 a PhD in 2024. 
When I graduated with a PhD in 


1132
01:07:11,572 --> 01:07:14,496
1984, forty years ago, I didn't 
have those things, right? 
 

1133
01:07:14,504 --> 01:07:18,560
So you guys have so much more 
exciting technology at your 
 

1134
01:07:18,568 --> 01:07:22,745
disposal, right? 
You have to figure out to solve 

1135
01:07:22,745 --> 01:07:27,165
 a world's problem of simulation
or base or whatever, I mean 
 

1136
01:07:27,173 --> 01:07:30,006
healthcare or, or, or whatever 
problem, right? 
 

1137
01:07:30,014 --> 01:07:34,260
Try to figure out the 
combination of quantum plus ML 


1138
01:07:34,268 --> 01:07:37,876
plus cloud plus whatever, right?
I mean, how can I solve the 
 

1139
01:07:37,884 --> 01:07:40,798
world's problem, right? 
You start with a problem that 
 

1140
01:07:40,806 --> 01:07:43,810
you're going to solve, right? 
And you have a choice. 
 

1141
01:07:43,818 --> 01:07:46,882
You could either work in a large
company and just join that and 


1142
01:07:46,890 --> 01:07:50,880
run on that treadmill and, and 
you'll be guaranteed $100,000– 


1143
01:07:50,888 --> 01:07:54,103
$200,000 salary. 
You have a home, you, you, you 


1144
01:07:54,111 --> 01:07:56,216
whatever, right? 
Or you are passionate. 
 

1145
01:07:56,224 --> 01:08:01,068
You see, I could take a risk do 
do something interesting. 
 

1146
01:08:01,076 --> 01:08:06,878
And, and my son who graduated 
from Berkeley, he actually chose

1147
01:08:06,878 --> 01:08:09,560

 the path of the startup, 
right? 

1148
01:08:09,560 --> 01:08:12,320
I mean, he, he, he could have 
 
joined many. 

1149
01:08:12,320 --> 01:08:14,440
He had a computer science degree

 from Berkeley. 

1150
01:08:14,440 --> 01:08:17,399
He could have interviewed, he 
 
had interviewed at all the large

1151
01:08:17,399 --> 01:08:20,120
companies in the Bay Area, but 

he chose to do a startup and 

1152
01:08:20,120 --> 01:08:22,880
he's working on a startup in the

 healthcare area. 

1153
01:08:23,279 --> 01:08:25,520
And I, I, I wish him all the 
 
luck, right? 

1154
01:08:25,520 --> 01:08:27,854
And, and so he has taken a risk.

 

1155
01:08:27,862 --> 01:08:30,880
He has taken a much less 
compensation in, during the 
 

1156
01:08:30,888 --> 01:08:34,943
years of a startup with the hope
of transforming the world in 
 

1157
01:08:34,951 --> 01:08:38,020
this healthcare startup called 
Sempre Health that he's doing 
 

1158
01:08:38,028 --> 01:08:40,505
right along with his wife, 
Anurati. 
 

1159
01:08:40,513 --> 01:08:45,229
So, so that is the the message. 
I would like to end it with 
 

1160
01:08:45,238 --> 01:08:46,724
that. 
Follow your passion. 
 

1161
01:08:46,732 --> 01:08:49,569
You have to take some risks in 
life, right? 
 

1162
01:08:49,578 --> 01:08:52,645
It is not easy, the life of 
startup, but if it is 
 

1163
01:08:52,653 --> 01:08:56,160
successful, you will transform 
the world and will transform 
 

1164
01:08:56,167 --> 01:08:57,800
your personal financial 
situation. 
 

1165
01:08:57,808 --> 01:09:00,965
But you shouldn't do a startup 
for the financial. 
 

1166
01:09:00,973 --> 01:09:04,761
You should do the startup 
because you really want to solve

1167
01:09:04,761 --> 01:09:09,048

 a really hard problem that the
world doesn't know how to solve 

1168
01:09:09,048 --> 01:09:11,482
 using the latest technologies 
that you have. 
 

1169
01:09:11,490 --> 01:09:15,622
And people have not figured out 
how to combine quantum plus AI 


1170
01:09:15,630 --> 01:09:18,215
plus HPC plus cloud plus IoT, 
right? 
 

1171
01:09:18,223 --> 01:09:21,160
And you are the first guy who 
did it, right? 
 

1172
01:09:21,167 --> 01:09:22,974
It's that interdisciplinary 
thing. 
 

1173
01:09:22,982 --> 01:09:27,185
The ability to assimilate is 
what is unique in the startup. 


1174
01:09:27,194 --> 01:09:31,702
So if take a problem that is 
really hard and solve it with 
 

1175
01:09:31,710 --> 01:09:35,920
the gadgets that you have today,
which Prith Banerjee in 1984 
 

1176
01:09:35,928 --> 01:09:40,618
did not have access to, you guys
in 2024 have so much more to 
 

1177
01:09:40,626 --> 01:09:45,219
work on. 
Now that that's great advice and

1178
01:09:45,219 --> 01:09:49,125

 that, yeah, the I agree the 
the passion you need to have, 

1179
01:09:49,125 --> 01:09:51,683
it's a 
 bit like doing a PhD. 
There's no point in doing a PhD 

1180
01:09:51,683 --> 01:09:53,700
 just for the sake of it, or 
you'll fail, or you'll. 
 

1181
01:09:53,707 --> 01:09:55,420
You can still attacking my 
voice, right? 
 

1182
01:09:55,428 --> 01:09:59,229
I I. 
Thank. 
 

1183
01:09:59,237 --> 01:10:01,624
You very much for inviting me 
for for the podcast. 
 

1184
01:10:01,632 --> 01:10:03,640
I really, really enjoyed it. 
Yeah. 
 

1185
01:10:03,648 --> 01:10:05,997
Thank you so much. 
This has been great. 
 

1186
01:10:06,005 --> 01:10:10,126
Ansys is very lucky to have you 
as their CTO, so thank you 
 

1187
01:10:10,134 --> 01:10:11,360
again. 
Thank you, Neil.