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 explain 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 talk to leading engineers from elite level sports like 5 00:00:12,840 --> 00:00:16,880 cycling and Formula One to some of the world's top academics to 6 00:00:16,880 --> 00:00:20,450 understand how fluid dynamics, machine learning and 7 00:00:20,450 --> 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 --> 00:00:33,800 helpful to you too. So sit back and enjoy this 11 00:00:33,800 --> 00:00:40,680 episode. Hi, and welcome back to the Neil 12 00:00:40,680 --> 00:00:43,960 Ashton Podcast. So today's guest is Joris Poort, 13 00:00:43,960 --> 00:00:47,080 who is the CEO and founder of Rescale. 14 00:00:47,320 --> 00:00:51,000 He's somebody who I wanted to talk to for a while as part of 15 00:00:51,160 --> 00:00:54,280 the opportunity to speak to people who have had that bold 16 00:00:54,280 --> 00:00:57,160 vision to, to create a start up, to create a new company. 17 00:00:57,240 --> 00:00:59,920 And I thought it was really interesting to try and learn 18 00:00:59,920 --> 00:01:03,000 from these people, you know, what motivated them to do in the 19 00:01:03,000 --> 00:01:06,120 1st place, what some of the lessons they've learnt in doing 20 00:01:06,120 --> 00:01:08,160 that. And it's particularly relevant, 21 00:01:08,480 --> 00:01:10,840 I guess, for the themes of of this podcast. 22 00:01:11,560 --> 00:01:15,920 Rescale is, is used by actually a lot of companies these days, 23 00:01:15,920 --> 00:01:17,760 you know, if they want to integrate more like high 24 00:01:17,760 --> 00:01:21,280 performance computing and and more recently more applied AI, 25 00:01:21,280 --> 00:01:24,360 you know, they've got hundreds of customers in this space, 26 00:01:24,360 --> 00:01:26,480 enterprise customers. So it's kind of interesting to 27 00:01:26,480 --> 00:01:30,240 speak to somebody who looks after that company and they have 28 00:01:30,240 --> 00:01:32,000 a good sense of what's coming next. 29 00:01:32,120 --> 00:01:34,600 It's been a topic also this podcast is looking at, you know, 30 00:01:34,600 --> 00:01:38,240 what's the future of AI and we we ended up talking for quite 31 00:01:38,240 --> 00:01:42,240 some time about the potential of agentic AI. 32 00:01:42,240 --> 00:01:45,680 This is something that, you know, myself and Joris turns out 33 00:01:45,680 --> 00:01:48,600 are quite aligned on in this could be quite transformative. 34 00:01:48,600 --> 00:01:51,360 You know, we previously mainly spoken about AI surrogates and 35 00:01:51,360 --> 00:01:54,640 we we also do talk about that. But I think the yeah, the the 36 00:01:54,640 --> 00:01:57,640 agentic AI and the potential this has for engineering was a 37 00:01:57,640 --> 00:01:59,840 really interesting discussion that comes probably in the 38 00:01:59,880 --> 00:02:03,520 second-half of of the chat. So it definitely tune in if you 39 00:02:03,520 --> 00:02:06,720 want to hear some interesting thoughts on that, given my 40 00:02:06,720 --> 00:02:09,440 background, having worked at AWS before and therefore being 41 00:02:09,440 --> 00:02:12,840 immersed in the cloud computing space, I always find that 42 00:02:12,840 --> 00:02:16,000 interesting as well. And how that has evolved. 43 00:02:16,080 --> 00:02:21,200 You know, Rescale was created like 2011, Oh, quite a long time 44 00:02:21,360 --> 00:02:24,990 before, I guess ways today where cloud is more mainstream and 45 00:02:24,990 --> 00:02:27,200 and accepted, you know, and Rescale's had some pretty 46 00:02:27,200 --> 00:02:31,960 impressive founders, sorry, fund funding from companies like 47 00:02:32,120 --> 00:02:35,440 people like Sam Altman and Jeff Bezos, Paul Graham. 48 00:02:35,440 --> 00:02:38,280 So talk a bit about that with the, you know, Y Combinator, 49 00:02:38,280 --> 00:02:41,320 NVIDIA, Microsoft. So, you know, it's, it's 50 00:02:41,360 --> 00:02:44,120 impressive feat for someone to create a company that's had 51 00:02:44,120 --> 00:02:45,840 hundreds of millions of dollars in funding. 52 00:02:46,280 --> 00:02:50,280 And so I hope that you learn a lot from some of his advice for 53 00:02:50,280 --> 00:02:52,600 for maybe one of you who is thinking about creating a, a 54 00:02:53,200 --> 00:02:56,240 start up yourself. I always talk about doing this 55 00:02:56,240 --> 00:02:58,600 and I'm probably just too risk-averse to do it. 56 00:02:58,600 --> 00:03:01,280 So, but don't listen to me, listen to Joris and hopefully 57 00:03:01,280 --> 00:03:02,700 you get some, some, some insights. 58 00:03:02,708 --> 00:03:04,149 So this was a wide-ranging discussion. 59 00:03:04,157 --> 00:03:07,682 You know, we probably could have taken it even deeper or in other 60 00:03:07,682 --> 00:03:10,882 areas, but at an hour and a half, I thought he was already 61 00:03:10,890 --> 00:03:13,730 taking a lot of his time. But yeah, I certainly learnt a 62 00:03:13,738 --> 00:03:16,810 lot from this discussion and have a renewed and even more 63 00:03:16,818 --> 00:03:19,897 sense of, you know, appreciation for what people like him do and 64 00:03:19,897 --> 00:03:22,760 push the boundaries and try and create new companies and ideas. 65 00:03:22,760 --> 00:03:24,240 So hopefully it's an inspiration 66 00:03:24,240 --> 00:03:27,680 for you listening as well. So sit back and enjoy this 67 00:03:27,680 --> 00:03:30,840 episode with Joris Poort. Thank you for for coming and 68 00:03:30,840 --> 00:03:33,800 doing this. You're actually on the quite 69 00:03:33,800 --> 00:03:37,560 high up the list of people I wanted to speak coming I guess 70 00:03:37,560 --> 00:03:41,520 from a cloud provider background before sort of seeing cloud. 71 00:03:41,800 --> 00:03:46,211 I think I joined AWS in 2020 and thinking, oh, this is quite 72 00:03:46,211 --> 00:03:49,120 novel, this is quite new. And then looking back and 73 00:03:49,120 --> 00:03:53,520 realized that Rescale was part of the Y Combinator in 2011. 74 00:03:54,640 --> 00:03:58,080 So back then, HPC in the cloud must have been even more 75 00:03:58,080 --> 00:04:00,480 radical, even more sort of crazy ideas. 76 00:04:00,480 --> 00:04:02,400 So maybe it's a starting point. 77 00:04:02,408 --> 00:04:06,570 I'd be interested to know like what was your pitch for that 78 00:04:06,578 --> 00:04:10,270 start up for for Rescale? Yeah, absolutely. 79 00:04:10,278 --> 00:04:13,080 So first of all, thanks for thanks for having me on. 80 00:04:13,088 --> 00:04:17,118 I think this is, you know a nice opportunity to to take some time 81 00:04:17,118 --> 00:04:21,635 and chat about the background. I think for HPC in the cloud, 82 00:04:21,642 --> 00:04:26,724 certainly when we started in 2011, it was a very new concept. 83 00:04:26,724 --> 00:04:29,400 I will be honest, when we 84 00:04:29,400 --> 00:04:32,920 founded the company, we thought we were late to market because 85 00:04:33,680 --> 00:04:35,160 cloud computing had already started. 86 00:04:35,400 --> 00:04:39,360 You had big data. And it seemed pretty obvious to 87 00:04:39,360 --> 00:04:42,120 me that there would be like a sort of a big compute company. 88 00:04:42,360 --> 00:04:46,920 And I actually looked for a company to join myself to say, 89 00:04:46,920 --> 00:04:51,000 hey, who who's doing this, solving this problem, right? 90 00:04:51,480 --> 00:04:53,880 And I had personally sort of experienced this problem before. 91 00:04:53,880 --> 00:04:57,600 So it seemed like we were late to market. 92 00:04:58,000 --> 00:05:00,200 Looking back now, we're quite early to market, right? 93 00:05:00,720 --> 00:05:03,560 So at that time there was definitely 0% HPC happening in 94 00:05:03,560 --> 00:05:05,480 cloud. But it's, you know, it's been a 95 00:05:05,480 --> 00:05:07,880 fun journey and things have, you know, changed over time. 96 00:05:08,640 --> 00:05:13,720 It was a tough process to get people excited about actually, 97 00:05:13,920 --> 00:05:16,394 you know, investing in this and, and sort of joining the team 98 00:05:16,394 --> 00:05:19,960 to pursue this. But you know, like most good 99 00:05:19,960 --> 00:05:23,823 start-ups, you can start with a a great idea and and and lots 100 00:05:23,823 --> 00:05:27,280 of effort and eventually you know you can make it. 101 00:05:28,000 --> 00:05:31,720 Yeah. So maybe taking a step before 102 00:05:31,720 --> 00:05:33,970 that, where had you been before? 103 00:05:33,978 --> 00:05:37,876 What gave you that original motivational idea even to 104 00:05:37,884 --> 00:05:41,320 overcome these problems and and create a start up, right. 105 00:05:41,328 --> 00:05:44,240 That's still a a leap of faith to leave a company to a start 106 00:05:44,248 --> 00:05:45,310 up. Yeah, absolutely. 107 00:05:45,318 --> 00:05:48,640 I think for for me, my background is quite technical. 108 00:05:48,648 --> 00:05:52,405 So I'd studied, I grew up doing a lot of computer science. 109 00:05:52,413 --> 00:05:54,620 I'd studied applied math, mechanical engineering, 110 00:05:54,628 --> 00:05:57,886 aeronautics, astronautics. I was sort of part of did a lot 111 00:05:57,886 --> 00:06:00,214 of work in the field of multi- disciplinary optimisation. 112 00:06:00,222 --> 00:06:04,454 I know you come from a strong CFD background, right. 113 00:06:04,462 --> 00:06:09,530 So by the professor I studied under actually was an 114 00:06:09,530 --> 00:06:13,495 aeroelasticity expert who was a disciple of Lucien Schmit from 115 00:06:13,503 --> 00:06:16,980 the sort of structure of the first person to kind of do 116 00:06:16,988 --> 00:06:19,490 structural optimization. So I had sort of a background 117 00:06:19,498 --> 00:06:22,699 there. And then I, I spent some time 118 00:06:22,707 --> 00:06:27,487 working at Boeing applying a lot of these different kind of 119 00:06:27,495 --> 00:06:29,821 tools. So a lot of software, a lot of 120 00:06:29,829 --> 00:06:34,395 math, a lot of different kind of physics calculations for the 787 121 00:06:34,395 --> 00:06:37,162 Dreamliner program. And the, the big challenge we 122 00:06:37,170 --> 00:06:41,215 had there was trying to solve, it's the first kind of fully 123 00:06:41,223 --> 00:06:44,200 carbon fiber airplane and the wing optimization, wing being 124 00:06:44,208 --> 00:06:47,065 kind of the most important part of an airplane. 125 00:06:47,073 --> 00:06:49,660 A lot of big technical challenges are the main 126 00:06:49,668 --> 00:06:52,456 difference being since it's carbon fiber, many more 127 00:06:52,464 --> 00:06:54,263 variables. And how do you sort of optimize 128 00:06:54,263 --> 00:06:57,277 this design? And so I had a background there 129 00:06:57,277 --> 00:07:00,440 kind of using different techniques in in sort of 130 00:07:00,448 --> 00:07:03,402 parameters trying to optimize what it would be the lightest 131 00:07:03,410 --> 00:07:05,178 weight wing design for the best performance. 132 00:07:05,186 --> 00:07:09,172 Long story short, took us many years, but eventually we got 133 00:07:09,180 --> 00:07:12,322 there and that's the wing if you've ever flown on 787, that's 134 00:07:12,322 --> 00:07:14,187 the that's the wing that's on there. 135 00:07:14,195 --> 00:07:17,256 So it's a very efficient airplane, right. 136 00:07:17,264 --> 00:07:20,900 But in order to do that and get to that answer, we we really 137 00:07:20,908 --> 00:07:23,494 have to leverage a lot of different computing 138 00:07:23,502 --> 00:07:25,735 capabilities. And this was more than 20 years 139 00:07:25,735 --> 00:07:27,229 ago. So there was no cloud computing 140 00:07:27,229 --> 00:07:29,763 yet. So it was really much more about 141 00:07:29,763 --> 00:07:33,104 how do we scale sort of a distributed systems problem from 142 00:07:33,104 --> 00:07:36,760 a software perspective inside of Boeing with, you know, 143 00:07:36,760 --> 00:07:38,828 different resources from different business units that we 144 00:07:38,828 --> 00:07:42,240 would sort of over the weekend be able to kind of gather a 145 00:07:42,240 --> 00:07:44,748 bunch of compute capacity service together from different 146 00:07:44,748 --> 00:07:47,790 teams and, and solve some of these larger scale problems. 147 00:07:47,798 --> 00:07:51,530 And eventually that that really got me into HPC because it was 148 00:07:51,538 --> 00:07:54,668 like in academia, I actually studied more how do you solve 149 00:07:54,676 --> 00:07:56,166 these equations more efficiently, right. 150 00:07:56,174 --> 00:07:59,730 So trying to combine aero- elasticity, try to kind of 151 00:07:59,738 --> 00:08:02,565 combine mechanical engineering physics with, with fluid 152 00:08:02,573 --> 00:08:04,166 dynamics and, and things like that. 153 00:08:04,174 --> 00:08:07,264 And with the ultimate goal of just, you know, more efficiently 154 00:08:07,264 --> 00:08:09,360 calculating all these different complex multiphysics 155 00:08:09,360 --> 00:08:11,760 responses. And at Boeing, we were 156 00:08:11,760 --> 00:08:14,200 implementing with this sort of multiphysics problem, but in a 157 00:08:14,200 --> 00:08:16,560 very different way than how academics kind of looked at the 158 00:08:16,560 --> 00:08:18,400 problem. It was much more about the, 159 00:08:18,880 --> 00:08:20,860 let's throw some more compute at this problem and like where 160 00:08:20,860 --> 00:08:23,277 are the real bottlenecks and like sort of how do you scale 161 00:08:23,277 --> 00:08:26,030 this? And from my own experience, I 162 00:08:26,038 --> 00:08:29,490 really enjoyed the ability to kind of gather a lot of these 163 00:08:29,498 --> 00:08:32,260 compute resources and just solve interesting problems faster. 164 00:08:32,268 --> 00:08:36,140 So that got me into this whole sort of category of HPC. 165 00:08:36,140 --> 00:08:38,240 I. Think one kind of insight as 166 00:08:38,240 --> 00:08:41,120 well was like, you know, if you look at kind of the people who 167 00:08:41,120 --> 00:08:45,040 were the best aerodynamicists at Boeing, for example, or the 168 00:08:45,040 --> 00:08:48,028 best at some of these like large physics computational 169 00:08:48,028 --> 00:08:52,226 problems, they became the really good at running HPC, right, 170 00:08:52,226 --> 00:08:54,800 Basically in practice, right. Like these are people at in 171 00:08:54,800 --> 00:08:56,200 industry sort of working on this. 172 00:08:56,200 --> 00:08:59,720 And I, I, I think I sort of had a background because of the 173 00:08:59,720 --> 00:09:02,280 software and the math to be able to solve those kinds of 174 00:09:02,280 --> 00:09:04,400 problems. But it's a really like a sort of 175 00:09:04,400 --> 00:09:08,160 untapped potential to be able to kind of unlock, you know, the, 176 00:09:08,200 --> 00:09:11,200 the possibility of, of leveraging really large scale 177 00:09:11,200 --> 00:09:13,000 compute for many different problems. 178 00:09:13,120 --> 00:09:16,520 And so that sort of. Led to hey. 179 00:09:17,280 --> 00:09:19,840 This is an area I'm like pretty passionate about. 180 00:09:20,160 --> 00:09:21,755 I also tried some other things. 181 00:09:21,763 --> 00:09:25,144 I was a management consultant for some period of time, short 182 00:09:25,152 --> 00:09:28,356 period of time, yeah. And so like I, I went to 183 00:09:28,364 --> 00:09:31,315 Business School. So as I looked at it, I saw a 184 00:09:31,323 --> 00:09:34,952 pretty broad set of different things you could do. 185 00:09:34,960 --> 00:09:38,518 But the thing that kept pulling me back was I, I did really 186 00:09:38,526 --> 00:09:40,872 think there was like a really interesting problem and and 187 00:09:40,880 --> 00:09:44,496 really impactful problem if we could solve this sort of large 188 00:09:44,504 --> 00:09:47,280 scale physics calculations for like engineers and scientists 189 00:09:47,288 --> 00:09:51,540 that really push the sort of boundaries forward, not only in 190 00:09:51,540 --> 00:09:54,816 a place like Boeing, but also in many other industries, right, 191 00:09:54,816 --> 00:09:57,280 in automotive and in life sciences, semiconductor. 192 00:09:57,288 --> 00:10:01,487 And so that seemed like a a worthwhile sort of mission to 193 00:10:01,495 --> 00:10:05,780 pursue. But I'm, I'm kind of intrigued 194 00:10:05,788 --> 00:10:09,279 the practicals of doing it because I often feel that a lot 195 00:10:09,279 --> 00:10:11,725 of engineers and maybe it's changed now, haven't got the 196 00:10:11,733 --> 00:10:15,262 mindset to create a start up that, you know, they, they think 197 00:10:15,262 --> 00:10:18,216 a little bit more linearly, you know, OK, I'm going to do 198 00:10:18,224 --> 00:10:19,418 engineering and make something better. 199 00:10:19,426 --> 00:10:23,162 Was it going to Business School or going to McKinsey that gave 200 00:10:23,170 --> 00:10:26,945 you a bit more confidence and awareness to go and do a start 201 00:10:26,953 --> 00:10:30,508 up? Yeah, I think, well, there's, 202 00:10:30,516 --> 00:10:33,780 there's many different things. I I think ultimately for like 203 00:10:33,788 --> 00:10:37,935 the best founders, they have like a few traits that are 204 00:10:37,943 --> 00:10:41,160 pretty common that together are very uncommon, right. 205 00:10:41,168 --> 00:10:45,440 So like it's see if I can sort of recall, I think Marc 206 00:10:45,448 --> 00:10:48,259 Andreessen sort of shared what his perspective on this is, 207 00:10:48,267 --> 00:10:51,071 which are like, like, I think a pretty good viewpoint. 208 00:10:51,079 --> 00:10:54,080 So what is you have to be very open and curious, right? 209 00:10:54,088 --> 00:10:56,514 So you have to be kind of willing to learn many different 210 00:10:56,514 --> 00:10:58,100 things. And so that's probably also 211 00:10:58,108 --> 00:11:01,148 pretty common with like people in academia, things like that. 212 00:11:01,156 --> 00:11:05,600 You also have to. Be be pretty. 213 00:11:05,608 --> 00:11:08,800 Willing to stick with something and work through a lot of 214 00:11:08,808 --> 00:11:10,395 challenges for a long period of time. 215 00:11:10,403 --> 00:11:13,900 And then there's this element of like he calls it 216 00:11:13,908 --> 00:11:17,110 disagreeableness, but it's like you sort of have to be stubborn 217 00:11:17,110 --> 00:11:19,845 enough, right to be able it's it's you know, starting a 218 00:11:19,853 --> 00:11:22,184 company is not the most like rational thing to do, right? 219 00:11:22,192 --> 00:11:26,600 Like no matter kind of which field you're in, it's, it's, 220 00:11:26,608 --> 00:11:30,812 it's really hard, right? And so it's like people do it 221 00:11:30,820 --> 00:11:34,188 because they kind of have to, not the people who just do it 222 00:11:34,196 --> 00:11:37,842 because they want to. It's, it's often for maybe not 223 00:11:37,850 --> 00:11:40,805 the, the sort of ideal reasons, right? 224 00:11:40,813 --> 00:11:44,120 But you'd be sure to have these different traits, you need a lot 225 00:11:44,120 --> 00:11:45,992 of like, I think it's, it's pretty risky. 226 00:11:46,000 --> 00:11:50,002 And so like, and a lot of people, you know, will tell you 227 00:11:50,002 --> 00:11:53,130 to do something different or why it's not going to work or 228 00:11:53,130 --> 00:11:55,076 why everybody's already thought of this idea from the 229 00:11:55,076 --> 00:11:55,880 founders. I know. 230 00:11:56,160 --> 00:11:58,360 I think it's just like, yeah, it's, it's a combination of 231 00:11:58,360 --> 00:11:59,840 these, all these different traits, right? 232 00:12:00,360 --> 00:12:03,720 And, and everyone is sort of unique, also uniquely flawed, 233 00:12:03,720 --> 00:12:09,000 probably in many ways. I certainly AM, but but it is 234 00:12:09,000 --> 00:12:12,560 hard to find. I would say like and out of all 235 00:12:12,560 --> 00:12:14,680 these capabilities, you have to be really smart of course and 236 00:12:14,680 --> 00:12:17,640 things like that. But like the probably the one in 237 00:12:17,640 --> 00:12:19,840 least supply is, is I think the courage, right? 238 00:12:19,920 --> 00:12:24,400 It's like sort of the willingness to kind of jump in 239 00:12:24,400 --> 00:12:26,720 and do it. There are times though, when 240 00:12:26,720 --> 00:12:29,120 entrepreneurship is much more popular, right? 241 00:12:29,120 --> 00:12:33,000 Like so when certain companies are taking off and there's a lot 242 00:12:33,000 --> 00:12:35,640 of funding and things look a little like a little easier, you 243 00:12:35,640 --> 00:12:37,160 get a lot of people jumping into the game. 244 00:12:37,600 --> 00:12:40,880 But most big companies are really successful ones are built 245 00:12:40,880 --> 00:12:42,520 over a very long period of time, right? 246 00:12:42,520 --> 00:12:43,760 And there's lots of ups and downs. 247 00:12:44,480 --> 00:12:46,735 Even if you look at the absolute most successful companies, 248 00:12:46,735 --> 00:12:49,080 there are really challenging periods. 249 00:12:49,088 --> 00:12:55,612 And I think I think you have to be willing to enable to kind of 250 00:12:55,612 --> 00:12:58,030 work through those, those difficult times, right. 251 00:12:58,038 --> 00:13:02,020 And yeah, it's maybe something from Jensen, I think I shared 252 00:13:02,028 --> 00:13:05,155 before where it's like, you know, like it, it's really the 253 00:13:05,163 --> 00:13:07,998 challenges that that form the character that allow you to to 254 00:13:08,006 --> 00:13:11,245 become like one of these kind of leaders of these companies. 255 00:13:11,253 --> 00:13:14,000 It's probably not just, yeah, just being smart or something 256 00:13:14,008 --> 00:13:15,980 like that. Yeah, what, what practically 257 00:13:15,988 --> 00:13:19,066 though were the steps? I'm always kind of intrigued. 258 00:13:19,074 --> 00:13:21,568 So the beginning, you have an idea. 259 00:13:21,576 --> 00:13:24,119 How did that form? Did you have an idea? 260 00:13:24,127 --> 00:13:27,776 And then you, you know, you go around to the various VCs to 261 00:13:27,784 --> 00:13:30,902 get to get funding. You had just a tiny idea. 262 00:13:30,910 --> 00:13:35,455 And then it evolved like what? How did this thing come to be, 263 00:13:35,463 --> 00:13:37,430 essentially? Yeah, there's a good book 264 00:13:37,438 --> 00:13:40,175 written on this from by Peter Thiel 0 to one, right? 265 00:13:40,183 --> 00:13:42,827 It's like like how do you start something from nothing? 266 00:13:42,835 --> 00:13:46,516 I think the how how it tactically works, right? 267 00:13:46,524 --> 00:13:51,062 It's it's for me, it was about sort of came to the conclusion 268 00:13:51,070 --> 00:13:53,230 you asked about this as well. Like what does Business School 269 00:13:53,238 --> 00:13:55,270 really teach you? I mean, I think the biggest 270 00:13:55,278 --> 00:13:58,022 lesson from Business School is maybe, you know, all these, at 271 00:13:58,030 --> 00:14:01,240 least the one I went to, all these different CEOs come 272 00:14:01,248 --> 00:14:02,960 through, right? And they give these talks. 273 00:14:02,968 --> 00:14:04,510 These are all very impressive people. 274 00:14:04,518 --> 00:14:08,930 But I think the one thing once you've seen enough of them, 275 00:14:08,938 --> 00:14:11,412 right, you do all these case studies and you learn all these 276 00:14:11,412 --> 00:14:13,670 things. It does give you the feeling 277 00:14:13,678 --> 00:14:18,110 that at least for me, you know, sort of you can really do 278 00:14:18,118 --> 00:14:19,862 anything, right? Like you hear these kind of 279 00:14:19,870 --> 00:14:22,738 stories, the same thing if you listen to like sort of founder 280 00:14:22,746 --> 00:14:25,516 stories. And I think you could kind of do 281 00:14:25,516 --> 00:14:27,664 anything you set your mind to do, right? 282 00:14:27,672 --> 00:14:30,688 And it's in some ways one of the best insurance policies, right? 283 00:14:30,688 --> 00:14:32,200 Because like, look, you get this 284 00:14:32,200 --> 00:14:34,800 Business School degree, you know, you can go get a job 285 00:14:34,800 --> 00:14:37,480 somewhere. Like I had sort of, you know, 286 00:14:37,480 --> 00:14:39,480 done this internship at McKinsey. 287 00:14:39,480 --> 00:14:44,640 I could go back there, right? And that gives you, you know, 288 00:14:44,640 --> 00:14:47,320 maybe sort of a floor of like, OK, now I can take a lot more 289 00:14:47,320 --> 00:14:49,920 risk right now. This was a time I had to, I had, 290 00:14:49,920 --> 00:14:52,760 I did not have a spouse that didn't have kids. 291 00:14:52,960 --> 00:14:57,240 I, I was actually at sort of a point in time, I think where I 292 00:14:57,240 --> 00:14:59,000 was able to take the most risk, right? 293 00:14:59,080 --> 00:15:01,920 And so I think that's sort of sets a good foundation to jump 294 00:15:01,920 --> 00:15:03,920 into to do it. I would recommend, right? 295 00:15:03,920 --> 00:15:06,000 Like you do have to kind of burn the boats, right? 296 00:15:06,000 --> 00:15:08,600 Like, so you can't be comparing yourself to your peers who are 297 00:15:08,600 --> 00:15:11,240 going to go work in finance or whatever and make a bunch of 298 00:15:11,240 --> 00:15:14,674 money or go work in academia and publish the most papers and 299 00:15:14,674 --> 00:15:16,640 the do most innovative kind of thinking. 300 00:15:16,960 --> 00:15:18,344 I have a lot of peers like that. 301 00:15:18,352 --> 00:15:21,520 And I think you have to be kind of get yourself to the point 302 00:15:21,528 --> 00:15:25,786 where like you, you really do want to start this company kind 303 00:15:25,786 --> 00:15:29,785 of no matter what, right? And then you kind of have to 304 00:15:29,793 --> 00:15:31,843 like most founders describe it as like, I would say that you 305 00:15:31,851 --> 00:15:34,098 kind of have to right, Like you just don't see any other way. 306 00:15:34,106 --> 00:15:35,895 And that's kind of the how I saw it. 307 00:15:35,903 --> 00:15:37,989 It was not to just for the purpose of starting company. 308 00:15:37,997 --> 00:15:40,720 It was like. To make the impact. 309 00:15:40,728 --> 00:15:44,062 That I thought would be possible like that I could make myself 310 00:15:44,070 --> 00:15:47,555 right and I saw kind of was possible at Boeing and then it's 311 00:15:47,555 --> 00:15:50,148 like OK there's many other industries there's many other 312 00:15:50,156 --> 00:15:53,200 engineers and scientists who are all like bottleneck by compute 313 00:15:53,208 --> 00:15:55,920 basically right and then there's this thing called cloud 314 00:15:55,928 --> 00:15:59,710 computing and like everybody's access to these and it seemed 315 00:15:59,718 --> 00:16:05,012 again I felt I was late right and so it seemed pretty obvious 316 00:16:05,012 --> 00:16:09,822 to get to go pursue that the. The tactical steps would be I 317 00:16:09,830 --> 00:16:12,023 decided to just move to Silicon Valley. 318 00:16:12,031 --> 00:16:17,715 So like, I came out Silicon Valley and I knew nothing right 319 00:16:17,715 --> 00:16:19,532 about. Silicon Valley, really like I'd, 320 00:16:19,532 --> 00:16:22,307 you know, like read some books, listened to some 321 00:16:22,307 --> 00:16:24,520 podcasts, probably weren't really, I don't even know. 322 00:16:24,528 --> 00:16:26,090 What they were. Called podcasts at that time, 323 00:16:26,098 --> 00:16:30,240 but the you know, Stanford had this like I think they still do 324 00:16:30,240 --> 00:16:32,120 this entrepreneurial thought leaders program. 325 00:16:32,128 --> 00:16:36,914 And so I remember just listening to these people come in and this 326 00:16:36,914 --> 00:16:39,242 is a different generation. So you talk about like, you 327 00:16:39,250 --> 00:16:41,384 know, I was listening to this stuff like like, like way back 328 00:16:41,392 --> 00:16:45,496 in the day, right? Like sort of 20, probably 2010 329 00:16:45,496 --> 00:16:48,366 maybe earlier to yeah, probably earlier. 330 00:16:48,374 --> 00:16:52,080 But they give you some inspiration and sort of a road 331 00:16:52,088 --> 00:16:55,340 map of of like, you know, how you can start a company. 332 00:16:55,348 --> 00:16:56,970 There is something special about Silicon Valley. 333 00:16:56,978 --> 00:16:59,970 So you have this very high concentration of founders, you 334 00:16:59,978 --> 00:17:05,109 know, engineers who want to build companies as sort of 335 00:17:05,117 --> 00:17:08,328 builders, investors. You have sort of a cultural 336 00:17:08,336 --> 00:17:11,760 appetite for risk, right? Like it's a, it's a very special 337 00:17:11,760 --> 00:17:15,000 place. And I do think if it's like my 338 00:17:15,008 --> 00:17:18,190 mentality is a little bit like, you know, if you're going to, if 339 00:17:18,190 --> 00:17:21,060 you're going to go do something like this, you might 340 00:17:21,060 --> 00:17:26,569 as well, you know, try to sort of play in the NBA, so to speak, 341 00:17:26,569 --> 00:17:28,760 right? Like so many places you can 342 00:17:28,760 --> 00:17:33,320 start a company, but I think to give yourself the best chances 343 00:17:33,320 --> 00:17:36,400 of success, it seems like moving to Silicon Valley would be a a 344 00:17:36,600 --> 00:17:39,640 good move, right? But I literally moved out 345 00:17:39,640 --> 00:17:43,680 without really, I had no family or you know, I had some 346 00:17:44,240 --> 00:17:46,480 classmates that also moved to the Bay Area. 347 00:17:46,520 --> 00:17:51,760 But I did not have any like like sort of real good reason to be 348 00:17:51,760 --> 00:17:53,720 there other than to kind of try to start a company. 349 00:17:54,640 --> 00:17:56,320 And I just started working at writing code. 350 00:17:56,640 --> 00:17:59,920 So I just started building the product in parallel. 351 00:17:59,920 --> 00:18:02,743 It was like, you know, trying to see, hey, you know, can we 352 00:18:02,743 --> 00:18:05,760 raise some money, etcetera. But it was just by myself and 353 00:18:05,760 --> 00:18:08,640 just writing some code. That's really how it started, 354 00:18:08,880 --> 00:18:10,240 right? And I think you mentioned 355 00:18:10,240 --> 00:18:13,680 earlier like sort of eventually led to Y Combinator, which is 356 00:18:13,680 --> 00:18:15,440 one of the incubators. So that that can really help a 357 00:18:15,440 --> 00:18:17,160 lot, right? Because they that can give you a 358 00:18:17,160 --> 00:18:20,880 pretty quick start and a very fast network of other founders 359 00:18:21,080 --> 00:18:24,680 and great mentors that I can get you off the ground. 360 00:18:25,480 --> 00:18:26,560 It's a little bit different than today. 361 00:18:26,560 --> 00:18:29,360 This is back in 2011, right? So at that time you got to 362 00:18:29,360 --> 00:18:32,080 remember it was like social, local, mobile or all the hot 363 00:18:32,080 --> 00:18:35,800 trends, right? And so like sort of working on, 364 00:18:37,080 --> 00:18:39,400 you know, a, a product that would do like. 365 00:18:39,400 --> 00:18:42,520 You. Know complex super computing for 366 00:18:42,520 --> 00:18:45,960 multidisciplinary optimization of physics where the aerospace 367 00:18:45,960 --> 00:18:49,680 market was definitely not the hottest idea, right yeah, yeah, 368 00:18:50,200 --> 00:18:55,000 but I think you know that what is nice is I do think the 369 00:18:55,000 --> 00:18:59,320 culture in Silicon Valley embraces this sort of anybody 370 00:18:59,320 --> 00:19:01,760 with an idea can come there right and you know ideas are 371 00:19:01,760 --> 00:19:04,240 cheap right so it's all about the building and the execution 372 00:19:04,240 --> 00:19:05,400 so. Yeah. 373 00:19:06,800 --> 00:19:10,320 It's in a very meritocratic place, like, unlike like many 374 00:19:10,320 --> 00:19:13,480 other sort of games, so to speak, in life. 375 00:19:13,560 --> 00:19:16,800 I feel like Silicon Valley is quite meritocratic, right? 376 00:19:17,400 --> 00:19:19,920 Where sort of anybody can be the next like Zuck. 377 00:19:20,840 --> 00:19:22,760 And so everybody kind of has to be nice to each other and help 378 00:19:22,760 --> 00:19:27,040 each other. And so, you know, it's yeah, it 379 00:19:27,040 --> 00:19:29,920 has a nice. I think that dynamic is really, 380 00:19:29,920 --> 00:19:31,804 there's a lot of paying it forward, like a lot of founders 381 00:19:31,804 --> 00:19:34,000 help each other. People are very accessible, 382 00:19:34,040 --> 00:19:37,680 right. So for me that was yeah, it was 383 00:19:37,680 --> 00:19:39,960 a it was a great journey. It was, it was pretty tough 384 00:19:39,960 --> 00:19:40,960 though. Like I like I was saying, I was 385 00:19:40,960 --> 00:19:43,240 pitching this company. Yeah, yeah. 386 00:19:43,360 --> 00:19:46,996 That, you know, the exact opposite of social, local, 387 00:19:46,996 --> 00:19:48,475 mobile. There's like literally that if 388 00:19:48,483 --> 00:19:51,139 you have the Venn diagrams, we would be the one that like does 389 00:19:51,139 --> 00:19:53,160 not overlap with anything that like investors were interested 390 00:19:53,168 --> 00:19:55,210 in at that time. But you know, everything goes 391 00:19:55,218 --> 00:19:58,297 through these waves, right? And so I do think eventually, 392 00:19:58,305 --> 00:20:01,980 like, you know, if you just have a good mission, you're on, 393 00:20:01,988 --> 00:20:03,367 right? Like you'll find some funding, 394 00:20:03,375 --> 00:20:06,608 right? Like it is a there's a lot of 395 00:20:06,616 --> 00:20:08,670 investors, right? And so ultimately you just need 396 00:20:08,670 --> 00:20:10,952 one to write a check. Yeah, yeah. 397 00:20:10,960 --> 00:20:13,374 So what was the first few years like? 398 00:20:13,382 --> 00:20:14,738 What were the some standout moments? 399 00:20:14,746 --> 00:20:19,560 Who were the was a turning point from a big investor or like in 400 00:20:19,560 --> 00:20:23,558 those early days, 2011, what when did you really feel that 401 00:20:23,558 --> 00:20:26,920 he was going to work out? Because I guess there were 402 00:20:26,928 --> 00:20:29,520 moments when you maybe thought, OK, I should just stop doing 403 00:20:29,528 --> 00:20:31,245 this and get a job somewhere else. 404 00:20:31,253 --> 00:20:34,660 Yeah, I mean, I think like, you know, what you call working out 405 00:20:34,660 --> 00:20:37,562 versus like success, these are all like your own definition, I 406 00:20:37,562 --> 00:20:40,530 would say, right, Like so and that definition for most people 407 00:20:40,530 --> 00:20:43,466 changes over time. I remember making a promise to 408 00:20:43,474 --> 00:20:46,482 my significant other at the time, we were not married yet 409 00:20:46,490 --> 00:20:48,975 that, you know, like I'll just do this thing. 410 00:20:48,983 --> 00:20:53,549 And she was working really hard and and, you know, like really 411 00:20:53,557 --> 00:20:55,577 grinding. And I was just kind of sitting 412 00:20:55,585 --> 00:20:58,052 in our, I was working really hard, but I was just writing 413 00:20:58,060 --> 00:21:00,220 software and not making any money and, and she was paying 414 00:21:00,228 --> 00:21:03,192 the rent. And so we sort of made this deal 415 00:21:03,192 --> 00:21:05,430 where like, OK, well, if you know, let's, you know, at what 416 00:21:05,438 --> 00:21:07,720 point are you going to say, how can I get a get a real job, 417 00:21:07,728 --> 00:21:09,305 right? And my parents were wondering 418 00:21:09,313 --> 00:21:11,630 the same thing. And we basically said, well, 419 00:21:11,638 --> 00:21:14,126 like, you know, let's give it two years. 420 00:21:14,134 --> 00:21:17,232 And if you can pay yourself a salary of something, right? 421 00:21:17,240 --> 00:21:21,176 Like enough to clear like the minimum medical benefits and 422 00:21:21,184 --> 00:21:23,960 things like that, then that's a win, right? 423 00:21:23,968 --> 00:21:25,460 Like that, then we can keep going basically. 424 00:21:25,468 --> 00:21:30,260 So that was kind of the. Deal I made with her and. 425 00:21:30,268 --> 00:21:35,932 I think that's you know, like, like, yeah, I would say really 426 00:21:35,940 --> 00:21:38,630 made it it it changes all the time. 427 00:21:38,638 --> 00:21:42,216 If I think if you have the ambition and sort of the mission 428 00:21:42,216 --> 00:21:44,525 that we're on, right? It's a it's a very big mission, 429 00:21:44,525 --> 00:21:47,368 right? And so, you know, I think I 430 00:21:47,376 --> 00:21:50,915 think we still have a lot of our work cut out for us even today, 431 00:21:50,915 --> 00:21:53,380 right? But but some big milestones, 432 00:21:53,388 --> 00:21:56,702 important lesson I think for founders is also like, you know, 433 00:21:56,702 --> 00:22:01,220 I think Paul Graham says this like, you know, companies don't 434 00:22:01,220 --> 00:22:05,840 die, founders give up, right? And you can't just keep going 435 00:22:05,848 --> 00:22:07,780 right now. There's there's a limit probably 436 00:22:07,780 --> 00:22:10,562 like you take market feedback, you're like, Hey, is this still, 437 00:22:10,562 --> 00:22:12,088 is this investable? Is this like smart? 438 00:22:12,096 --> 00:22:13,629 Like did you learn some new things? 439 00:22:13,637 --> 00:22:16,526 But if you're kind of iterating quickly, you're learning a lot 440 00:22:16,534 --> 00:22:18,532 of things. You're adapting your company to 441 00:22:18,532 --> 00:22:20,760 like meet the market, so to speak, right? 442 00:22:20,768 --> 00:22:24,765 Kind of get this product market fit and then any set of like 443 00:22:24,773 --> 00:22:27,800 really smart people that, that work really hard together and 444 00:22:27,808 --> 00:22:31,414 sort of have these attributes they, they will build success, 445 00:22:31,422 --> 00:22:33,614 right? Like, and so I think it is in 446 00:22:33,622 --> 00:22:36,026 that way an amazing place where you can kind of pursue, pursue 447 00:22:36,034 --> 00:22:39,860 what you want to do. For us, a big milestone was 448 00:22:39,868 --> 00:22:42,856 first check. So the first investor is always 449 00:22:42,856 --> 00:22:46,245 very special for us. That was, you know, we'd gone 450 00:22:46,253 --> 00:22:49,547 around like pitched all these like VCs kept getting turned 451 00:22:49,555 --> 00:22:51,343 down. And at this point we were doing 452 00:22:51,343 --> 00:22:53,882 Y Combinator. And you know, Paul Graham teed 453 00:22:53,890 --> 00:22:57,580 up like one of his buddies. It's like, OK, just give him the 454 00:22:57,580 --> 00:22:59,700 pitch. Here's the thing, I followed 455 00:22:59,708 --> 00:23:03,270 exactly the playbook that he gave me and you know, we didn't, 456 00:23:03,270 --> 00:23:07,350 we didn't know what price, what, you know, amount of the 457 00:23:07,350 --> 00:23:09,295 company we should give up for, for whatever investment around. 458 00:23:09,295 --> 00:23:10,520 We're just trying to get some 459 00:23:10,520 --> 00:23:14,320 investor right. And it was like, OK, well, like, 460 00:23:14,520 --> 00:23:17,680 like, how about these terms or whatever based on guidance from 461 00:23:17,680 --> 00:23:19,600 Paul Graham. And, you know, eventually he 462 00:23:19,600 --> 00:23:22,880 goes through a bunch of things and still don't invest. 463 00:23:23,280 --> 00:23:25,920 So I come back to Paul and I'm just like, man, you know, we did 464 00:23:25,920 --> 00:23:28,680 everything you said. This is like the, you know, 50th 465 00:23:28,680 --> 00:23:34,080 time and, you know, we didn't get the check and but but you 466 00:23:34,080 --> 00:23:36,480 know, I did follow like all the things that you said, right? 467 00:23:36,480 --> 00:23:37,880 And so like, why don't you just invest? 468 00:23:37,880 --> 00:23:41,640 And he's like, sure, he whips out his checkbook and writes a 469 00:23:41,640 --> 00:23:44,440 check. And, and I think that was very 470 00:23:44,440 --> 00:23:46,600 special. Not really about the actual 471 00:23:46,600 --> 00:23:50,440 dollar amount invested. It was much more about his, you 472 00:23:50,440 --> 00:23:53,000 know, when somebody kind of makes a bet on you like that, 473 00:23:53,080 --> 00:23:54,920 right? And it's like real money out of 474 00:23:54,920 --> 00:23:58,160 his pocket. It, you know, means so much to 475 00:23:58,160 --> 00:24:01,080 you that I'm sure you've seen this with maybe like a thesis 476 00:24:01,080 --> 00:24:03,640 advisor or other people who've been in your career in the past. 477 00:24:04,160 --> 00:24:07,000 That that is a a really special moment, right? 478 00:24:07,560 --> 00:24:10,520 And I think that certainly gave me a lot of confidence to kind 479 00:24:10,520 --> 00:24:14,080 of just keep, keep going. And eventually we, we got lots 480 00:24:14,080 --> 00:24:16,520 of great investors, but that was that was a special one. 481 00:24:16,520 --> 00:24:19,240 Another one was, you know, we, we'd started working with some 482 00:24:19,240 --> 00:24:20,920 early customers like we couldn't raise money. 483 00:24:20,920 --> 00:24:24,520 So we just like built the product and started to get who 484 00:24:24,520 --> 00:24:27,840 wants to go run like large scale CFD in the cloud, right? 485 00:24:29,200 --> 00:24:33,480 And it's like, you know, there's like Boeing, Airbus, etcetera. 486 00:24:33,480 --> 00:24:35,800 At that time there were these really small space companies, 487 00:24:35,800 --> 00:24:37,880 we're talking about like 50 person companies, right? 488 00:24:39,440 --> 00:24:43,680 And we went to, there were basically only two private space 489 00:24:43,680 --> 00:24:46,800 companies at the time. We went to both of them and 490 00:24:47,000 --> 00:24:48,680 because of my aerospace background, I've known some 491 00:24:48,680 --> 00:24:51,920 people that working there. So we're able to kind of chat 492 00:24:51,920 --> 00:24:55,480 with them and eventually we sort of got them on the hook that 493 00:24:55,480 --> 00:25:00,760 they they would do this right. And that that was also very 494 00:25:00,760 --> 00:25:03,000 special because I think that's when you kind of. 495 00:25:03,520 --> 00:25:05,840 Could see hey. Like, and they're making a bet 496 00:25:05,840 --> 00:25:07,640 on you, right? And, and, you know, when the 497 00:25:07,640 --> 00:25:09,920 company is like one or two people, it's very personal, 498 00:25:10,080 --> 00:25:12,880 right? So when these with these first 499 00:25:12,880 --> 00:25:15,440 few customers, like at this point, I had my Co founder, 500 00:25:15,440 --> 00:25:17,840 who's my old boss from Boeing, Adam McKenzie join. 501 00:25:17,840 --> 00:25:21,480 And, you know, we sit there in the room with some of these like 502 00:25:21,480 --> 00:25:24,468 aerodynamicists and they're just like this concept of putting 503 00:25:24,468 --> 00:25:27,720 it in the cloud is, is pretty foreign. 504 00:25:27,760 --> 00:25:31,684 And remember, like web services, So AWS had just started and 505 00:25:31,684 --> 00:25:34,800 web services are built for like almost the exact opposite of 506 00:25:34,800 --> 00:25:36,800 HPC, right? Like it, it's like very loosely 507 00:25:36,800 --> 00:25:40,480 coupled, very, you know, you kind of have to write your 508 00:25:40,480 --> 00:25:42,640 software very tolerant to failure, etcetera. 509 00:25:43,920 --> 00:25:46,280 The actual performance is like not great. 510 00:25:47,080 --> 00:25:52,160 And so it, you know, but you could kind of see where this was 511 00:25:52,160 --> 00:25:55,400 going. And so we pitched these 512 00:25:55,400 --> 00:25:58,370 companies and I think that was a big breakthrough when when 513 00:25:58,370 --> 00:26:00,296 they were like, yes, we'll buy the software, We'll pay 514 00:26:00,296 --> 00:26:01,680 something for this basically right. 515 00:26:01,688 --> 00:26:05,080 And we'll buy the software and again, that was many iterations 516 00:26:05,080 --> 00:26:07,180 to get there. But once you get that 517 00:26:07,188 --> 00:26:09,628 breakthrough, I mean, that was very exciting, right? 518 00:26:09,636 --> 00:26:14,200 And and those little wins along the way, I think it's like at no 519 00:26:14,200 --> 00:26:16,640 point, I think the entire journey I felt like I've really 520 00:26:16,640 --> 00:26:17,990 made it. Yeah. 521 00:26:17,998 --> 00:26:22,615 But you know, like I, I think those are very material in those 522 00:26:22,615 --> 00:26:25,120 early days, right? Like like just just getting off 523 00:26:25,120 --> 00:26:27,820 the ground. I think many great ideas just 524 00:26:27,828 --> 00:26:31,712 don't have enough energy on execution put behind them, 525 00:26:31,720 --> 00:26:33,522 right? And, and even to this day, 526 00:26:33,530 --> 00:26:36,498 right, like I still have people come to me and be like, well, I 527 00:26:36,498 --> 00:26:38,688 mean, we're doing much more than just HPC these days. 528 00:26:38,696 --> 00:26:41,482 But you get some, you know, you still have this sort of old 529 00:26:41,490 --> 00:26:44,820 network of HPC people, you know, I both know who we're talking 530 00:26:44,828 --> 00:26:46,664 about and. You know there'll be. 531 00:26:46,672 --> 00:26:48,710 People who say you get both messages, right? 532 00:26:48,718 --> 00:26:51,824 There were people at that time who were like, you get these 533 00:26:51,832 --> 00:26:54,630 Silicon Valley investors who were like, what do you mean like 534 00:26:54,630 --> 00:26:57,252 HPC, Like it already exists, It's cloud computing and it's 535 00:26:57,260 --> 00:26:59,804 AWS, right? So you have that class of people 536 00:26:59,804 --> 00:27:02,807 which are basically sort of the, yeah, it's like it's 537 00:27:02,807 --> 00:27:04,255 already been done. Like why are you wasting your 538 00:27:04,263 --> 00:27:05,920 time on this? And then you have the other 539 00:27:05,928 --> 00:27:07,328 class, which are the HPC people, right? 540 00:27:07,336 --> 00:27:10,280 Like it's super computing. And these these folks are like, 541 00:27:10,280 --> 00:27:12,454 this is absolutely impossible, like never possible. 542 00:27:12,462 --> 00:27:14,962 All that hardware is not good enough, like etcetera. 543 00:27:14,970 --> 00:27:18,908 And you have an entire ecosystem that these days it's a little 544 00:27:18,916 --> 00:27:21,302 bit different, right? But remember then it was like 545 00:27:21,310 --> 00:27:24,474 nobody was really doing this and nobody believed it was really 546 00:27:24,482 --> 00:27:28,340 possible. And then you have people who 547 00:27:28,348 --> 00:27:33,000 say, well, I had that idea. Right, as if having the idea is 548 00:27:33,000 --> 00:27:35,782 an important thing. Yeah, I think the execution is 549 00:27:35,790 --> 00:27:37,927 the the big thing that really matters. 550 00:27:37,935 --> 00:27:40,315 And that's where like those traits of a good founder, I 551 00:27:40,323 --> 00:27:42,934 think you, you, you know, willing to persevere. 552 00:27:42,942 --> 00:27:45,440 If you look at most companies, if you're just willing to kind 553 00:27:45,448 --> 00:27:49,330 of go for more than a decade, most of the companies are very 554 00:27:49,338 --> 00:27:53,045 successful, right? But it's not easy, right? 555 00:27:53,053 --> 00:27:56,270 So like I, I think that's also like, it's not for everybody. 556 00:27:56,278 --> 00:28:00,240 Well, and, and I was thinking that you were probably really 557 00:28:00,248 --> 00:28:04,965 facing an uphill battle because, you know, if I look at from a 558 00:28:04,973 --> 00:28:07,784 technology point of view, you know, the hyperscalers took 559 00:28:07,792 --> 00:28:13,264 quite a long time to really have, for example, the 560 00:28:13,272 --> 00:28:16,577 interconnect. You, you, you know, like when I 561 00:28:16,577 --> 00:28:21,110 joined AWS in what 2020, things like EFA or they, they were 562 00:28:21,118 --> 00:28:24,356 quite new. So you were, did you feel that 563 00:28:24,364 --> 00:28:28,160 was part of the challenge before that, the hyperscalers that I 564 00:28:28,168 --> 00:28:32,630 guess you know, you were simplifying the use of when 565 00:28:32,638 --> 00:28:38,756 moving as quick as you wanted in terms of like technology? 566 00:28:38,764 --> 00:28:43,784 Yeah, very good question because like that was the remember we 567 00:28:43,792 --> 00:28:48,040 started in 2011, right. So there were some some days in 568 00:28:48,040 --> 00:28:53,664 the wilderness there, but. With cloud computing was kind of 569 00:28:53,664 --> 00:28:54,720 really gaining traction, right? 570 00:28:54,840 --> 00:28:56,960 And even back then, you got to remember people were way 571 00:28:56,960 --> 00:28:59,880 underestimating these markets by like X, right? 572 00:28:59,960 --> 00:29:03,240 Like from where it is today? It shows how difficult it is. 573 00:29:03,240 --> 00:29:04,880 Even when you see something happening, it's like very 574 00:29:04,880 --> 00:29:07,080 difficult to forecast, right? We could talk about AI later, 575 00:29:07,360 --> 00:29:12,120 yeah. Early on, a big one was always 576 00:29:12,120 --> 00:29:14,640 like, well, like, like why do I need Rescale, right? 577 00:29:14,640 --> 00:29:18,640 Like I could just go to a cloud company and get this myself, 578 00:29:18,640 --> 00:29:20,760 right? And so and. 579 00:29:21,080 --> 00:29:23,760 A lot of people perceived even to this day, probably people 580 00:29:23,760 --> 00:29:28,120 perceived the cloud companies that are very close partners as 581 00:29:28,120 --> 00:29:31,120 like competitors in this market and that's sort of the wrong 582 00:29:31,120 --> 00:29:34,960 framing. The reality was I would be going 583 00:29:34,960 --> 00:29:37,560 in at that time. It was going to be Andy Jassy 584 00:29:38,160 --> 00:29:42,320 and begging him to build InfiniBand networking and 585 00:29:42,320 --> 00:29:45,000 begging him to build like Specs that. 586 00:29:45,880 --> 00:29:48,680 Would do really well for this market, right? 587 00:29:48,720 --> 00:29:52,400 And the answer often wasn't necessarily no, but it was like, 588 00:29:52,400 --> 00:29:56,600 well, an incremental dollar spent today, you know, I'm 589 00:29:56,600 --> 00:29:59,760 better off just building more, let's call it for simplification 590 00:29:59,760 --> 00:30:02,680 commodity compute like because there's more market share to 591 00:30:02,680 --> 00:30:06,560 grab there versus this like specialized compute, everything 592 00:30:06,560 --> 00:30:11,040 costs X more, workloads are much more volatile, etcetera, 593 00:30:11,040 --> 00:30:12,880 etcetera, etcetera, right. So there were sort of many 594 00:30:12,880 --> 00:30:15,120 reasons to not do it. There's also like some real 595 00:30:15,120 --> 00:30:18,440 difficult networking challenges like security challenges. 596 00:30:18,440 --> 00:30:20,280 So the the way you build like a cloud. 597 00:30:20,280 --> 00:30:23,880 Service as, as you know very well is a little bit different 598 00:30:23,880 --> 00:30:27,320 than the HPC system, right. And so at that time, the 599 00:30:27,320 --> 00:30:32,302 maturity wasn't there to easily sort of pursue this market, 600 00:30:32,302 --> 00:30:33,705 but. But from first. 601 00:30:33,713 --> 00:30:35,512 Principles, it seems like a really a lost opportunity, 602 00:30:35,520 --> 00:30:38,644 right, because if you kind of look at it, it's like, well, you 603 00:30:38,644 --> 00:30:40,800 know, commodity compute great. Like the whole thing about AWS 604 00:30:40,808 --> 00:30:43,320 was, hey, it was built for scaling e-commerce peak loads 605 00:30:43,328 --> 00:30:46,112 during the like, Black Friday or something, right? 606 00:30:46,120 --> 00:30:50,042 And but you know, like the rest of the time and even today, like 607 00:30:50,042 --> 00:30:53,302 if you just look at it wasn't the money makers at AWS, it's 608 00:30:53,310 --> 00:30:57,198 like the simple, it's just like EC2, S3 like basic stuff, right? 609 00:30:57,198 --> 00:30:59,160 And most people consume just 610 00:30:59,160 --> 00:31:03,160 that what and super computing and high performance computing 611 00:31:03,160 --> 00:31:05,400 was always the case is that, you know, these systems are super 612 00:31:05,400 --> 00:31:07,560 expensive. So time sharing them only the 613 00:31:07,560 --> 00:31:10,080 biggest companies in the world or the biggest labs could afford 614 00:31:10,080 --> 00:31:12,000 to kind of build these systems, right? 615 00:31:12,080 --> 00:31:15,040 And it makes a lot of sense to share those systems, right, 616 00:31:15,040 --> 00:31:16,200 because they're so expensive, right? 617 00:31:16,200 --> 00:31:18,840 So share the CapEx investment, the sort of timeshare of these 618 00:31:18,840 --> 00:31:20,600 systems, so? Used to be like mainframe time 619 00:31:20,600 --> 00:31:22,920 sharing, but like with cloud this becomes so much easier. 620 00:31:23,440 --> 00:31:28,400 And so like it really did seem like that was A at that time 621 00:31:28,800 --> 00:31:31,717 like a a lost opportunity and as Rescale we did make a 622 00:31:31,717 --> 00:31:33,800 conscious choice. It's like we are not at least at 623 00:31:33,800 --> 00:31:35,520 that time, right. We're not going to go get into 624 00:31:35,520 --> 00:31:39,360 this CapEx game like the venture capital dollars are not made. 625 00:31:39,400 --> 00:31:40,640 You can see that today with these. 626 00:31:40,840 --> 00:31:43,720 AI companies in the infrastructure side are not 627 00:31:43,720 --> 00:31:46,920 really made for, you know, making big CapEx compute 628 00:31:46,920 --> 00:31:50,760 investments, right. And so, yeah, we would be trying 629 00:31:50,760 --> 00:31:52,416 to work really closely with the. 630 00:31:52,424 --> 00:31:54,664 Cloud providers to actually build the right infrastructure 631 00:31:54,672 --> 00:31:56,612 for our customers. But you could kind of see where 632 00:31:56,612 --> 00:31:58,384 it's going, right? Like at that time it was to me 633 00:31:58,392 --> 00:32:02,050 again, I thought it was late. It was pretty obvious that like 634 00:32:02,050 --> 00:32:04,914 this is a great market. Opportunity, right, But like 635 00:32:04,922 --> 00:32:08,542 yeah, yeah, I think of the eyes of say somebody like Andy Jassy 636 00:32:08,542 --> 00:32:11,195 or you know, Satya was running advanced computing at that time 637 00:32:11,195 --> 00:32:13,848 in Azure. I think they made some smart 638 00:32:13,856 --> 00:32:18,186 bets too, I think. I think they, yeah, HPC is still 639 00:32:18,186 --> 00:32:21,465 this like small sub segment that seems very difficult 640 00:32:21,465 --> 00:32:24,616 relative to all of computing. So all these things are all 641 00:32:24,624 --> 00:32:26,489 about your frame, right? Like if you come from the HPC 642 00:32:26,497 --> 00:32:28,187 frame, this seems like pretty obvious I think. 643 00:32:28,195 --> 00:32:31,560 If you come from the, you're running like the fastest 644 00:32:31,568 --> 00:32:34,598 growing, most successful cloud company in the world, like, you 645 00:32:34,598 --> 00:32:36,393 know, there's many bets you could be making. 646 00:32:36,401 --> 00:32:39,178 And yeah, it took a while to get this one off the ground even 647 00:32:39,186 --> 00:32:40,854 today. And as far as I understand, 648 00:32:40,862 --> 00:32:43,620 there's no InfiniBand networking at at AWS. 649 00:32:43,628 --> 00:32:47,160 No, that's, that's true. And we should say as well, I 650 00:32:47,168 --> 00:32:50,342 guess that when we use the term HPC and I guess the relevant 651 00:32:50,350 --> 00:32:53,415 first discussing this is, I mean you can argue, but I would say 652 00:32:53,423 --> 00:32:57,082 from a hardware point of view, what AI training needs today has 653 00:32:57,082 --> 00:33:00,570 many, many, many of the same components as HPC. 654 00:33:00,578 --> 00:33:05,240 There's some subtleties, but in terms of a low latency network, 655 00:33:05,240 --> 00:33:08,900 that's basically what AI training needs as well. 656 00:33:08,908 --> 00:33:12,718 So interestingly, like the reason that we're talking about 657 00:33:12,718 --> 00:33:15,918 prior in those early days, it was only really for loosely 658 00:33:15,926 --> 00:33:18,968 coupled, wasn't it, from like a hardcore performance point of 659 00:33:18,976 --> 00:33:22,114 view. So my question is, do you think 660 00:33:22,114 --> 00:33:26,160 those early days almost made it harder to convince people when 661 00:33:26,168 --> 00:33:30,440 there was actually the hardware? Are you still getting people who 662 00:33:30,440 --> 00:33:34,144 are like it's it's slow running on the cloud or there's 663 00:33:34,144 --> 00:33:37,660 a because they they're still just thinking back to the 664 00:33:37,660 --> 00:33:41,280 mid-2010s when maybe the hardware wasn't there. 665 00:33:41,288 --> 00:33:44,202 Do you know what I mean? Are people lazy to be thinking 666 00:33:44,210 --> 00:33:46,864 of what's current? Yeah, and and like you and I are 667 00:33:46,864 --> 00:33:50,772 probably on the same side of the table on this, but, but I 668 00:33:50,772 --> 00:33:55,385 think if I had to steelman, like what's the sort of reason 669 00:33:55,385 --> 00:33:58,572 to not not go to cloud, right? Like a lot of people bring up 670 00:33:58,580 --> 00:34:00,080 costs and we can talk about that. 671 00:34:00,088 --> 00:34:05,235 But I think if I'm sort of put my hat on as one of our 672 00:34:05,243 --> 00:34:08,719 customers, right like like a manufacturing company it seems 673 00:34:08,728 --> 00:34:12,415 if you're. Certainly starting a new company 674 00:34:12,415 --> 00:34:15,172 like like really obvious. To to sort. 675 00:34:15,179 --> 00:34:18,320 Of buy this compute as a service as opposed to start building 676 00:34:18,328 --> 00:34:19,857 your own data centers etcetera, right? 677 00:34:19,864 --> 00:34:22,487 Like unless there is some sort of reason that's your 678 00:34:22,495 --> 00:34:25,670 competitive advantage, right? I can tell you Boeing thought it 679 00:34:25,670 --> 00:34:26,440 was their competitive advantage. 680 00:34:26,440 --> 00:34:27,920 They probably still take that to this day. 681 00:34:27,920 --> 00:34:31,560 And yeah, there's, there are some I would say like edge cases 682 00:34:31,560 --> 00:34:34,400 where it can make sense. But in general, right, you want 683 00:34:34,400 --> 00:34:38,440 to kind of a provider. Like any major cloud provider 684 00:34:38,480 --> 00:34:41,960 has way better economics, has way more efficient sort of 685 00:34:41,960 --> 00:34:44,840 scale. The better supply chain they 686 00:34:44,840 --> 00:34:47,880 have like better hardware, faster, they are able to refresh 687 00:34:47,880 --> 00:34:50,880 and resell your old hardware if you want to rotate to new SKUs 688 00:34:50,880 --> 00:34:53,440 and things like that. So there's so many benefits if 689 00:34:53,440 --> 00:34:55,400 you just. Look at solving the real problem 690 00:34:55,400 --> 00:34:57,360 you're trying to solve as a manufacturer, say running a 691 00:34:57,360 --> 00:35:00,920 simulation faster or designing a vehicle faster. 692 00:35:01,400 --> 00:35:03,120 That's the real problem you're trying to solve, right? 693 00:35:03,720 --> 00:35:07,600 The problem I think in HPC. Is there's an entire industry 694 00:35:07,600 --> 00:35:10,080 that's been set up with a different framing, which is like 695 00:35:10,080 --> 00:35:14,800 I'm, I'm actually here to most efficiently invest in CapEx, run 696 00:35:14,800 --> 00:35:18,360 the infrastructure into the ground, try to efficiently use 697 00:35:18,360 --> 00:35:21,360 this infrastructure, you know, depreciate it in a really smart 698 00:35:21,360 --> 00:35:22,840 way. But it's all about this sort of 699 00:35:22,840 --> 00:35:25,720 IT/TCO sort of frame, right? And. 700 00:35:26,240 --> 00:35:30,160 Purely on that frame, maybe sometimes on-prem is competitive 701 00:35:30,160 --> 00:35:32,760 with cloud, right? But but if you even value a 702 00:35:32,760 --> 00:35:35,480 little bit like actual performance, which is in the 703 00:35:35,480 --> 00:35:38,200 name of the industry, right, like high performance computing, 704 00:35:38,320 --> 00:35:41,280 the simple sort of to me. The concept is, is like, well, 705 00:35:41,280 --> 00:35:42,800 who do you want to bear the CapEx? 706 00:35:42,800 --> 00:35:46,400 Do you want it to be some really low cost of capital, very big 707 00:35:46,400 --> 00:35:48,400 cloud infrastructure player, right? 708 00:35:48,520 --> 00:35:52,160 Should that be you, right, as as a manufacturer and especially 709 00:35:52,160 --> 00:35:55,900 with the sort of rotating to new infrastructure, there's a lot 710 00:35:55,900 --> 00:35:59,280 of complexity, right. So typical organization might be 711 00:35:59,280 --> 00:36:03,280 running, you say 50 different physics codes. 712 00:36:03,760 --> 00:36:06,080 Right. Each of these have different 713 00:36:06,080 --> 00:36:08,360 algorithms, some of them have different sub algorithms within 714 00:36:08,360 --> 00:36:10,320 those algorithms, right, Different ways of running them. 715 00:36:10,360 --> 00:36:13,000 And so there's like 10s of thousands of combinations and 716 00:36:13,000 --> 00:36:15,040 way to run this. Then they can all be compiled 717 00:36:15,040 --> 00:36:16,640 differently. They have different optimal 718 00:36:16,640 --> 00:36:17,640 architectures. Right. 719 00:36:17,720 --> 00:36:21,120 And it's sort of like, and then you run them sporadically at 720 00:36:21,120 --> 00:36:23,560 different types like so if you're a automotive company, 721 00:36:23,560 --> 00:36:25,920 there's a big part of the design cycle in detailed design. 722 00:36:25,920 --> 00:36:27,280 You know, this is happening a lot. 723 00:36:27,720 --> 00:36:31,400 There's other times where you may not be running so much and 724 00:36:31,400 --> 00:36:36,600 so it is pretty tough. I think for the average 725 00:36:36,720 --> 00:36:42,880 enterprise like you know, consumer of HPC to sort of say, 726 00:36:43,160 --> 00:36:45,440 you know, it makes sense to kind of run your own HPC. 727 00:36:45,480 --> 00:36:49,680 Now if you look at the market today, right, it's, it's about 728 00:36:49,680 --> 00:36:54,960 20% cloud and say 80% on-prem, which which you know, I don't 729 00:36:54,960 --> 00:36:56,880 know, blows my mind still today, right. 730 00:36:58,520 --> 00:37:00,920 I think there is a, you know, it's a big. 731 00:37:01,280 --> 00:37:03,080 Switching costs takes a long time. 732 00:37:03,080 --> 00:37:04,800 If you're running and operating data centers, you're maybe 733 00:37:04,800 --> 00:37:07,280 depreciating that you know if. It's a CFO making a decision 734 00:37:07,280 --> 00:37:11,440 maybe six years, right? And so even if you said 100% to 735 00:37:11,440 --> 00:37:14,040 cloud tomorrow, it's it still takes a long time, right? 736 00:37:15,640 --> 00:37:18,120 And there's a lot of complexities that I think are 737 00:37:18,600 --> 00:37:20,520 glossed over. So I think a big reason you 738 00:37:20,520 --> 00:37:24,880 didn't see it take off earlier was a lot of this complexity. 739 00:37:24,880 --> 00:37:29,160 It is actually very hard to run, say, a fluid dynamics code. 740 00:37:29,880 --> 00:37:33,880 On an HPC system, very scalably, very reliably and continuously 741 00:37:33,880 --> 00:37:36,600 keep the codes updated. And like when you run different 742 00:37:36,600 --> 00:37:38,480 algorithms, they scale differently and sort of like 743 00:37:38,480 --> 00:37:39,880 what's the right cluster size so. 744 00:37:40,400 --> 00:37:43,520 There's so much that goes into that. 745 00:37:43,520 --> 00:37:48,120 That is, it's still, you know, it's a specialized field right 746 00:37:48,280 --> 00:37:51,080 now, at Rescale we write software to simplify all that. 747 00:37:51,080 --> 00:37:52,800 Right. And try to sort of abstract away 748 00:37:52,800 --> 00:37:55,880 all this complexity, but it's quite complex and if you really 749 00:37:55,880 --> 00:37:58,760 dig into it. If you ask the people who are on 750 00:37:58,760 --> 00:38:00,884 the sort of still running on-prem systems, they would 751 00:38:00,884 --> 00:38:04,420 often tell you the biggest reason why is, is sort of this 752 00:38:04,420 --> 00:38:07,204 it is hard and you have to, you have to be willing to kind of 753 00:38:07,204 --> 00:38:11,040 reinvent yourself, right? So like, just imagine if you're 754 00:38:11,080 --> 00:38:15,720 a high performance computing administrator for like 1 of the 755 00:38:15,720 --> 00:38:17,960 top three aerospace companies in the world, right? 756 00:38:19,120 --> 00:38:21,800 All you've done last 20 years is basically figure out how to 757 00:38:21,800 --> 00:38:25,280 procure like the right infrastructure, you know, 18 758 00:38:25,280 --> 00:38:28,160 months of like. Procurement processes get these 759 00:38:28,160 --> 00:38:31,200 systems tested up and running service, you know your different 760 00:38:31,200 --> 00:38:33,320 engineering customers, internals to your organization. 761 00:38:33,320 --> 00:38:35,200 With cloud, it's a completely different paradigm. 762 00:38:35,200 --> 00:38:37,680 Like that entire paradigm is like sort of hardware up 763 00:38:37,680 --> 00:38:39,640 thinking, right? Which is like it sort of starts 764 00:38:39,640 --> 00:38:42,320 with the processor and then you build out the systems. 765 00:38:43,360 --> 00:38:44,880 I think the right way to think about. 766 00:38:44,880 --> 00:38:48,720 What is the purpose of HPC is ultimately to serve the 767 00:38:48,720 --> 00:38:50,392 workloads with good performance. 768 00:38:50,400 --> 00:38:53,268 And that's more like workload down thinking, right? 769 00:38:53,276 --> 00:38:56,136 So like it's really just about OK, like for this workload, what 770 00:38:56,136 --> 00:38:58,260 is actually the right infrastructure, what's the right 771 00:38:58,260 --> 00:38:59,840 price, what's the right performance, etcetera. 772 00:38:59,848 --> 00:39:02,060 Like how are you going to optimize that problem? 773 00:39:02,068 --> 00:39:05,236 And with an elastic system, you're going to solve that 774 00:39:05,244 --> 00:39:07,195 problem much more elegantly, right? 775 00:39:07,203 --> 00:39:15,382 So I would say that's a very long answer, but I think it's 776 00:39:15,390 --> 00:39:19,512 still to this day, you know, there's people who believe. 777 00:39:19,520 --> 00:39:23,023 Like HPC should be done on-prem Yeah I. 778 00:39:23,031 --> 00:39:26,565 I see that my big observation is always just timelines that 779 00:39:26,573 --> 00:39:30,938 everything, as you say, just takes so much longer than people 780 00:39:30,938 --> 00:39:34,541 realize. So, you know, when I started, I 781 00:39:34,541 --> 00:39:39,440 remember and this is This is why I'm still amazed by you 782 00:39:39,440 --> 00:39:43,629 having these ideas nine years earlier that people were like 783 00:39:43,629 --> 00:39:47,081 completely anti cloud. A lot of people were anti cloud. 784 00:39:47,081 --> 00:39:48,920 And then you know, just as the 785 00:39:48,920 --> 00:39:53,800 time as I was leaving, I remembered that it wasn't, are 786 00:39:53,800 --> 00:39:56,320 we going to do that? It was like how, how we could, 787 00:39:56,440 --> 00:40:01,600 oh, no, that's not true. It was probably hybrid, hybrid 788 00:40:01,600 --> 00:40:04,280 and we are going to do some of it on, but we'll still keep 789 00:40:04,800 --> 00:40:06,000 some. And that depended whether it's 790 00:40:06,000 --> 00:40:08,920 to start upon enterprise. But I definitely saw a shift. 791 00:40:08,920 --> 00:40:14,968 But then I thought that means the time then to do a PoC to 792 00:40:14,968 --> 00:40:18,620 then roll it out to change like so. 793 00:40:18,628 --> 00:40:22,485 It's probably, I guess you'd agree there's probably even more 794 00:40:22,485 --> 00:40:25,500 upside in cloud to come because it takes so long to 795 00:40:25,500 --> 00:40:28,750 convince people and then for them to change that we may not 796 00:40:28,750 --> 00:40:31,040 see that consequence for another few years. 797 00:40:31,048 --> 00:40:33,965 Yeah, I think that's right. I do think that, you know, 798 00:40:33,973 --> 00:40:37,794 Rescale, we're really focused on solving kind of like what is the 799 00:40:37,794 --> 00:40:40,344 real problem the customer's trying to solve, right? 800 00:40:40,352 --> 00:40:44,030 And so often it's again, using this example of a manufacturer 801 00:40:44,038 --> 00:40:48,016 might be an engineer doing sort of fluid dynamic simulation. 802 00:40:48,024 --> 00:40:51,744 They just want to run that simulation as efficiently as 803 00:40:51,752 --> 00:40:53,700 possible and like all this other stuff. 804 00:40:53,708 --> 00:40:56,600 And so they're like all the IT stakeholders trying to kind of 805 00:40:56,608 --> 00:41:00,010 serve that user ideal, right? You kind of work from that 806 00:41:00,018 --> 00:41:04,075 customer backwards. I think you really want to solve 807 00:41:04,075 --> 00:41:05,860 this. Problem more from first 808 00:41:05,868 --> 00:41:09,215 principles as opposed to. Sort of you have all this 809 00:41:09,223 --> 00:41:11,494 baggage of these like on-prem systems, hybrid, all this stuff, 810 00:41:11,494 --> 00:41:13,461 right? And our view is sort of like, 811 00:41:13,469 --> 00:41:16,846 well, if we can give you a service that is the optimal 812 00:41:16,854 --> 00:41:18,857 router, right? So they can real time say, OK, 813 00:41:18,865 --> 00:41:20,550 you're going to run this fluid dynamics code. 814 00:41:20,558 --> 00:41:22,932 You know, we understand some of the metadata, right? 815 00:41:22,940 --> 00:41:27,080 So we say, hey, it has like, you know, 500,000 cells or whatever, 816 00:41:27,080 --> 00:41:29,111 right? And say you're running 817 00:41:29,111 --> 00:41:31,680 STAR-CCM+. We know that version of STAR- 818 00:41:31,680 --> 00:41:35,520 CCM+, how well it scales. We we know like with that kind 819 00:41:35,520 --> 00:41:37,840 of cell model, like what's the size cluster that should go on, 820 00:41:37,880 --> 00:41:38,800 right. And then you go through this 821 00:41:38,800 --> 00:41:42,160 sort of what we've built as a compute recommendation engine. 822 00:41:42,160 --> 00:41:45,320 So you actually so say, OK, you can plug in your own on-prem 823 00:41:45,320 --> 00:41:46,760 resource. So that's one of the fixed 824 00:41:46,760 --> 00:41:48,160 resources you could choose from. 825 00:41:48,168 --> 00:41:50,940 It comes at some price layer internal sort of cost to 826 00:41:50,948 --> 00:41:52,704 operate. Maybe they have your cloud 827 00:41:52,712 --> 00:41:54,442 resources, right. And then, you know, just have 828 00:41:54,450 --> 00:41:56,860 one, you have all the major cloud providers, you have 829 00:41:56,860 --> 00:41:58,912 neoclouds, maybe have some special supercomputing 830 00:41:58,912 --> 00:42:00,640 relationships with universities, etcetera. 831 00:42:01,000 --> 00:42:04,520 In this whole network, hundreds of architecture choices, right? 832 00:42:04,640 --> 00:42:06,320 All of these had a little nuances, right? 833 00:42:06,320 --> 00:42:07,720 Oh, but that's an Intel processor. 834 00:42:07,720 --> 00:42:11,520 So you need to use Intel MPI or like hey here this like version 835 00:42:11,520 --> 00:42:13,160 of this code. Like needs to be compiled 836 00:42:13,160 --> 00:42:17,480 differently whatever. So it's a, it's a we want an 837 00:42:17,480 --> 00:42:20,040 analysis in this little data, but how many combinations are 838 00:42:20,040 --> 00:42:22,040 there is like more than 50 million combinations. 839 00:42:23,160 --> 00:42:24,840 So it's like, how do you solve that problem? 840 00:42:24,840 --> 00:42:28,960 Well, you basically want to go through some sort of algorithm 841 00:42:28,960 --> 00:42:31,520 and we. Recommendation system was kind 842 00:42:31,520 --> 00:42:34,080 of our solution to this problem to automatically kind of figure 843 00:42:34,080 --> 00:42:35,302 figure out and route that right. 844 00:42:35,310 --> 00:42:39,058 And that's now you have the service for this user for any 845 00:42:39,066 --> 00:42:42,230 workload. That you know, you can kind of 846 00:42:42,238 --> 00:42:45,260 use both the all the metadata and all the knowledge Rescale 847 00:42:45,268 --> 00:42:48,068 has from running many of these similar workloads before, but 848 00:42:48,076 --> 00:42:51,641 you can also leverage any of the resources you want. 849 00:42:51,649 --> 00:42:54,108 In the way you as an administrator can say, hey, 850 00:42:54,116 --> 00:42:56,940 like, I first want to fill up my own system or whatever, right? 851 00:42:56,948 --> 00:42:59,715 And that is, I think that's the right way to solve this problem, 852 00:42:59,715 --> 00:43:03,780 right? And it is, it is a pretty 853 00:43:03,788 --> 00:43:06,678 technical way, right? I think there's also like 854 00:43:06,686 --> 00:43:08,400 usability of this is, is super important. 855 00:43:08,408 --> 00:43:11,666 So for us, how this shows up is literally like you open up 856 00:43:11,674 --> 00:43:14,072 Rescale, you drag in your like input file, right? 857 00:43:14,080 --> 00:43:17,555 Or if you upload the input file and you and you select the 858 00:43:17,563 --> 00:43:19,472 software, you press run, right? It is that simple. 859 00:43:19,480 --> 00:43:22,176 But in that is this highly complex Configurator which can 860 00:43:22,184 --> 00:43:24,612 kind of optimize and solve these problems for you. 861 00:43:24,620 --> 00:43:26,343 And, you know, I, I think that's a. 862 00:43:26,351 --> 00:43:30,662 Great way to solve this problem. But like you said, like there's 863 00:43:30,662 --> 00:43:32,475 a, there's a lot of practical challenges, right? 864 00:43:32,483 --> 00:43:35,680 Like, so that's the technical person and he says, OK, like 865 00:43:35,688 --> 00:43:38,655 this is a great solution, right? There's a lot of practical 866 00:43:38,663 --> 00:43:41,238 things in enterprise software. And like go to market and like 867 00:43:41,246 --> 00:43:43,814 how do you use customers like transition and etcetera. 868 00:43:43,822 --> 00:43:47,818 So that are still challenges for customers today. 869 00:43:47,826 --> 00:43:51,594 But I think increasingly, like you said, it's easier and 870 00:43:51,602 --> 00:43:52,880 easier. Like this PoC process you 871 00:43:52,888 --> 00:43:55,312 described, you got to remember, like if you're buying on-prem, 872 00:43:55,320 --> 00:43:59,045 you're probably going through like a sort of 9 to 12 month 873 00:43:59,053 --> 00:44:01,498 like procurement process at a minimum, right? 874 00:44:01,506 --> 00:44:03,965 For many companies, it's closer to 18 months. 875 00:44:03,973 --> 00:44:06,340 You go through all this evaluation, then your system is 876 00:44:06,340 --> 00:44:09,740 finally live 18 months later. And that's just not like, you 877 00:44:09,748 --> 00:44:13,240 know, if you're an engineer trying to run this. 878 00:44:13,248 --> 00:44:15,832 Fluid dynamic simulation, you can't do that right? 879 00:44:15,840 --> 00:44:17,907 Like you got to solve the problem in a different way, 880 00:44:17,915 --> 00:44:20,448 right? And I do think today if you go 881 00:44:20,456 --> 00:44:23,382 to Rescale, right, and go through this process, you can do 882 00:44:23,382 --> 00:44:25,712 this all in like 5 minutes, right? 883 00:44:25,720 --> 00:44:28,240 But there are practical challenges. 884 00:44:28,248 --> 00:44:30,720 So if you think again, think of a like a large enterprise 885 00:44:30,728 --> 00:44:33,950 organization, there's a lot of new ways of thinking that's sort 886 00:44:33,950 --> 00:44:36,860 of required. And it comes back I think to 887 00:44:36,868 --> 00:44:38,703 like the courage, right? Like are you, are you willing 888 00:44:38,711 --> 00:44:41,448 to? Say, hey, like we got to do it a 889 00:44:41,448 --> 00:44:44,040 different way and I know I'm going to go through lots of 890 00:44:44,048 --> 00:44:46,862 challenges, etcetera. But and I think increasingly a 891 00:44:46,870 --> 00:44:51,010 lot of companies are seeing so much success that it is hard to 892 00:44:51,010 --> 00:44:53,704 argue the other way, right? Like if I had to argue where if 893 00:44:53,704 --> 00:44:55,947 you're using cloud, like if I go to on-prem. 894 00:44:55,955 --> 00:44:58,880 Yeah. There's a few reasons, but for 895 00:44:58,888 --> 00:45:01,200 most organizations this would make very little sense. 896 00:45:01,208 --> 00:45:05,186 And I guess now it seems, and I didn't predict it, but it seems 897 00:45:05,186 --> 00:45:08,506 to be happening that there's even more because of you see, 898 00:45:08,506 --> 00:45:12,004 sort of, neoclouds, which I guess has been really because of 899 00:45:12,004 --> 00:45:14,320 AI. There seems to be even more 900 00:45:14,920 --> 00:45:17,560 competition to the big cloud providers. 901 00:45:17,720 --> 00:45:20,960 I guess there's more options, but even more reason to have a 902 00:45:20,960 --> 00:45:23,760 single place to route that through. 903 00:45:23,800 --> 00:45:26,000 Because if you have an if you have accounts with cloud 904 00:45:26,000 --> 00:45:30,606 providers A, B, C, D and E, for that enterprise to build out 905 00:45:30,606 --> 00:45:33,160 all themselves. I mean it can be done. 906 00:45:33,160 --> 00:45:37,840 But it's a lot of work. Does it make it even more 907 00:45:37,840 --> 00:45:40,960 valuable to have like a go between it than it was when 908 00:45:40,960 --> 00:45:43,960 there was mainly just AWS as the biggest one? 909 00:45:45,560 --> 00:45:47,320 Yeah, I think. Now of course we can, we can 910 00:45:47,320 --> 00:45:48,880 optimize within a single cloud, right? 911 00:45:48,880 --> 00:45:51,800 So like if a customer really loves Microsoft or Amazon or 912 00:45:51,800 --> 00:45:54,920 Google like we can, we can certainly do that and and you 913 00:45:54,920 --> 00:45:56,720 sort of solve the problem at a smaller scale. 914 00:45:56,760 --> 00:46:00,120 I think there's this margins piece, there's the speed, right? 915 00:46:00,120 --> 00:46:01,834 That's why I think. You see the neoclouds as well, 916 00:46:01,834 --> 00:46:03,960 like they're just able to stand up a cluster much faster, 917 00:46:03,960 --> 00:46:06,160 like it is the bigger organization is it is hard, 918 00:46:06,160 --> 00:46:06,800 right? And it's pretty. 919 00:46:06,880 --> 00:46:09,680 Interesting to see Microsoft that they outsource a lot of 920 00:46:09,680 --> 00:46:13,560 their GPU compute to CoreWeave, right? 921 00:46:13,560 --> 00:46:16,728 So public information, you know, there's, there's a few 922 00:46:16,728 --> 00:46:19,000 dynamics I think why like 1 is. 923 00:46:19,008 --> 00:46:22,772 Sort of just pure GPU scarcity. So like they just have them, not 924 00:46:22,772 --> 00:46:25,014 only GPUs, they also have the power infrastructure, right? 925 00:46:25,022 --> 00:46:28,020 Like I'm sitting in a data center working to power them and 926 00:46:28,020 --> 00:46:30,102 cool them. So there's sort of that timing 927 00:46:30,110 --> 00:46:31,820 effect, but there is a setting up. 928 00:46:31,828 --> 00:46:33,930 Those systems are like HPC systems, right? 929 00:46:33,938 --> 00:46:38,056 And that is very different than building public cloud resources 930 00:46:38,056 --> 00:46:40,560 historically. Now, if you only have like one 931 00:46:40,568 --> 00:46:43,050 or two customers, you can build these bespoke systems, right? 932 00:46:43,058 --> 00:46:45,962 Like, like what cloud provider to get at is serving thousands 933 00:46:45,970 --> 00:46:48,272 or 10s of thousands, millions of customers, right? 934 00:46:48,280 --> 00:46:52,132 And so there's a challenge for a neocloud is how do you scale 935 00:46:52,140 --> 00:46:53,866 that? There's also, I think an element 936 00:46:53,866 --> 00:46:55,325 of, but the speed really matters. 937 00:46:55,333 --> 00:46:58,060 Like I, I think right now in AI, it's just all about speed, 938 00:46:58,068 --> 00:47:00,894 right? And so people are willing to do 939 00:47:00,894 --> 00:47:03,444 absolutely crazy. Thanks for speed, right? 940 00:47:03,452 --> 00:47:07,669 And it's because of the sort of ultimate end-use case like this 941 00:47:07,669 --> 00:47:11,072 sort of AI war between all the big tech companies is fueling 942 00:47:11,080 --> 00:47:14,770 all of that, right. And and if you're just a few 943 00:47:14,778 --> 00:47:17,320 months faster, I'm training a large language model that can 944 00:47:17,328 --> 00:47:19,307 make a big. Difference right in in the long 945 00:47:19,307 --> 00:47:22,250 term game here. And so it's, it's, it's 946 00:47:22,258 --> 00:47:23,720 interesting the bottlenecks, right? 947 00:47:23,728 --> 00:47:25,452 Like if you, I don't know if you've been following this, but 948 00:47:25,452 --> 00:47:27,424 there's like, it's all these neoclouds of course. 949 00:47:27,432 --> 00:47:30,516 But there's also these practical problems people are running 950 00:47:30,524 --> 00:47:33,295 into, like there's not enough electricians to, like, you know, 951 00:47:33,295 --> 00:47:38,170 build data centres, right? And there's like, you know, 952 00:47:38,178 --> 00:47:41,425 like, like a meta doesn't have enough like buildings. 953 00:47:41,433 --> 00:47:44,360 So they'd like operating data centers in tents, right? 954 00:47:44,368 --> 00:47:48,200 Like, and then there's like this really efficient cooling systems 955 00:47:48,200 --> 00:47:52,544 built by NVIDIA that like run super efficiently. 956 00:47:52,552 --> 00:47:57,264 But then like xAI is, is is like literally like, like using 957 00:47:57,272 --> 00:48:02,064 generators to do sort of cool water and pump it into the 958 00:48:02,072 --> 00:48:04,046 cooling system in the Super inefficient way. 959 00:48:04,054 --> 00:48:06,095 But they, they got their training model up much faster 960 00:48:06,103 --> 00:48:09,082 than anybody else, right? And so like, yeah, like running 961 00:48:09,082 --> 00:48:12,242 fast matters here. And that's probably where, to be 962 00:48:12,242 --> 00:48:15,492 fair, the, there is a difference between the HPC for 963 00:48:15,492 --> 00:48:20,134 CAE and the, and the HPC for AI because it's probably true. 964 00:48:20,142 --> 00:48:24,506 If if you're a very large AI company where you need so much 965 00:48:24,514 --> 00:48:27,598 compute, but it's basically just for you, you could probably 966 00:48:27,606 --> 00:48:30,109 argue that it might be actually better for you to build it 967 00:48:30,117 --> 00:48:31,704 yourself. If that's your core 968 00:48:31,712 --> 00:48:34,188 differentiator is to get that training model out. 969 00:48:34,196 --> 00:48:37,388 Two months later, you can kind of see like, why would we go and 970 00:48:37,388 --> 00:48:39,335 wait on a cloud? We can just do it ourselves. 971 00:48:39,343 --> 00:48:40,880 We, we have the same type of GPU. 972 00:48:40,888 --> 00:48:43,374 We have 100,000 of them. That's it. 973 00:48:43,382 --> 00:48:47,252 Then it's I can see the logic but for most manufacturing 974 00:48:47,260 --> 00:48:50,288 companies or aerospace companies, they don't have 975 00:48:50,296 --> 00:48:53,464 100,000 GPUs to train one model right? 976 00:48:53,472 --> 00:48:58,254 It's a different world. So I guess it's different needs 977 00:48:58,254 --> 00:49:03,228 as you say. And probably they don't have 150 978 00:49:03,228 --> 00:49:05,392 different applications, legacy applications like CAE does. 979 00:49:05,400 --> 00:49:07,472 I guess the AI world is something newer. 980 00:49:07,480 --> 00:49:11,950 Yeah, I do think like you know, AI is is changing a lot and you 981 00:49:11,950 --> 00:49:14,280 know the CAE engineering world as well, right. 982 00:49:14,288 --> 00:49:16,688 And, and we could talk more about that, but like you spend a 983 00:49:16,688 --> 00:49:18,312 lot of time talking about the HPC layer. 984 00:49:18,320 --> 00:49:21,814 I think the way, if you sort of Fast forward to Rescale today, 985 00:49:21,822 --> 00:49:23,574 right? Like the way the way we look at 986 00:49:23,574 --> 00:49:27,078 it today is there's the HPC layer or the compute layer, but 987 00:49:27,078 --> 00:49:30,504 you have the data layer, super important, right? 988 00:49:30,512 --> 00:49:36,086 Like, so, so one of the big argue like, OK, why do these big 989 00:49:36,086 --> 00:49:38,259 training models? Why, why are these then sort of 990 00:49:38,259 --> 00:49:39,806 on-prem is like a natural question, right? 991 00:49:39,814 --> 00:49:42,920 Like, but like if you're sort of the downloading entire web and 992 00:49:42,928 --> 00:49:46,664 then like training on that, that's a very heavy data gravity 993 00:49:46,664 --> 00:49:50,619 kind of situation, right? And of course you want that data 994 00:49:50,619 --> 00:49:52,760 as close as possible to the compute. 995 00:49:52,768 --> 00:49:54,367 So there's always like memory implications. 996 00:49:54,375 --> 00:49:57,280 There's there's how these, you know, all the detailed 997 00:49:57,288 --> 00:49:59,168 architecture really matters, but also it's like, how do you serve 998 00:49:59,168 --> 00:50:02,644 all these things? So that's one, I would say class 999 00:50:02,644 --> 00:50:06,558 of HPC problems. But data starts, you know, it's 1000 00:50:06,558 --> 00:50:10,878 true that some of this is only for these large LLM providers. 1001 00:50:10,886 --> 00:50:14,664 But you know, I think everybody like if I speak with our 1002 00:50:14,672 --> 00:50:16,920 customers, they're all building AI models themselves as well. 1003 00:50:16,928 --> 00:50:18,768 And these are much smaller scale. 1004 00:50:18,776 --> 00:50:21,774 But ultimately they need efficient infrastructure to 1005 00:50:21,782 --> 00:50:24,774 train their own data as well. Yeah. 1006 00:50:24,782 --> 00:50:27,396 And then obviously inference is a big one too, right? 1007 00:50:27,404 --> 00:50:29,988 And like running inference efficiently is going to get 1008 00:50:29,996 --> 00:50:34,085 really important right today, you know, if you if you sort of 1009 00:50:34,085 --> 00:50:35,895 breakdown the economics. So like Silicon Valley is 1010 00:50:35,903 --> 00:50:38,040 funding a lot of very cheap services, right? 1011 00:50:38,048 --> 00:50:39,980 Why? Because everybody wants to win 1012 00:50:39,988 --> 00:50:42,400 the AI platform war, so to speak, right? 1013 00:50:42,408 --> 00:50:45,738 And so people are willing to lose a lot of money today on 1014 00:50:45,746 --> 00:50:48,540 inference, right? Just to win the customer, right? 1015 00:50:48,540 --> 00:50:50,760 But over over time, all these 1016 00:50:50,760 --> 00:50:54,360 economics need to be figured out and things do need to be run 1017 00:50:54,360 --> 00:50:57,920 actually efficiently, right? And so a manufacturer, because 1018 00:50:58,480 --> 00:51:01,280 here say a Tier 1 manufacturer and automotive, you're kind of 1019 00:51:01,280 --> 00:51:04,680 running your it's like a 5–10% margin business often, right? 1020 00:51:04,760 --> 00:51:08,080 And so it's like you already have to run really efficiently. 1021 00:51:08,080 --> 00:51:10,040 Why that's why IT focus so much on TCO. 1022 00:51:10,040 --> 00:51:12,760 That's that's why it's like, hey, I can run like the system 1023 00:51:12,760 --> 00:51:15,600 for like 10 years and divide it, you know, like it all makes 1024 00:51:15,600 --> 00:51:18,280 sense to me. But I think if you think about 1025 00:51:18,280 --> 00:51:20,400 the engineering. Problems you want to solve if 1026 00:51:20,400 --> 00:51:24,320 you're a little bit, you know for thinking of like OK, you 1027 00:51:24,320 --> 00:51:26,720 know how to get out of being a 10% margin business as a. 1028 00:51:27,480 --> 00:51:31,000 Tier one supplier, right? It's I think the way out is 1029 00:51:31,000 --> 00:51:35,800 actually using the latest technologies, which is, you 1030 00:51:35,800 --> 00:51:38,320 know, there's, there's a compute layer, but especially your 1031 00:51:38,320 --> 00:51:40,035 data, say you're a seat manufacturer. 1032 00:51:40,043 --> 00:51:42,592 I know how to like do seat design really well. 1033 00:51:42,600 --> 00:51:44,756 And I might be really good at crash simulation, right? 1034 00:51:44,764 --> 00:51:47,466 And so and so have all this crash simulation data. 1035 00:51:47,474 --> 00:51:50,200 If I'm like one of the top three companies in seat manufacturing, 1036 00:51:50,200 --> 00:51:53,125 I probably have some of the best crash simulation data for 1037 00:51:53,125 --> 00:51:56,198 seats out of anybody, right? If you can build what we call 1038 00:51:56,206 --> 00:51:58,360 some of these like AI physics models around that, you can 1039 00:51:58,368 --> 00:52:00,904 build a lot of the intelligence into a much more compressed 1040 00:52:00,912 --> 00:52:02,700 timeline, right? So now instead of like trying to 1041 00:52:02,700 --> 00:52:04,668 speed up the algorithm, you're actually like changing the 1042 00:52:04,676 --> 00:52:07,080 algorithm and going to from deterministic to probabilistic. 1043 00:52:07,080 --> 00:52:09,600 But then maybe I can serve my 1044 00:52:09,600 --> 00:52:12,560 OEM like way faster, like not a little bit faster. 1045 00:52:12,680 --> 00:52:14,640 Like I can actually say, hey, like I, I think I could do that 1046 00:52:14,640 --> 00:52:19,000 seat with like 98% confidence and I could give you that answer 1047 00:52:19,080 --> 00:52:23,720 in maybe an hour instead of like 2 months of analysis, right? 1048 00:52:24,680 --> 00:52:28,760 And so this is, that's, I think where the future's headed. 1049 00:52:28,760 --> 00:52:32,440 And then like you have, that's on the simulation side. 1050 00:52:32,920 --> 00:52:36,000 Then there's also this, you know, we're working with 1051 00:52:36,000 --> 00:52:38,240 customers to build sort of engineering agents. 1052 00:52:38,240 --> 00:52:39,160 Right. It's like. 1053 00:52:39,240 --> 00:52:42,520 You, you kind of get this capability to this, you're 1054 00:52:42,520 --> 00:52:44,840 actually starting for the simulation engineers. 1055 00:52:44,840 --> 00:52:48,680 They have this this automotive company empowering them to just 1056 00:52:48,680 --> 00:52:52,320 do their job like way faster. So all the mundane tasks you're 1057 00:52:52,320 --> 00:52:54,000 starting to sort of automate, right? 1058 00:52:54,680 --> 00:52:57,440 And that's I think every function and every industry 1059 00:52:57,440 --> 00:53:00,160 frankly is going to go through that sort of process, right. 1060 00:53:01,160 --> 00:53:04,480 The nice thing in engineering is there is a lot of data and 1061 00:53:04,480 --> 00:53:07,080 there's a lot of like intelligence already built into 1062 00:53:07,080 --> 00:53:11,440 all these processes, etcetera. And so it's Rescale with our 1063 00:53:11,440 --> 00:53:12,920 customers. We're in a unique position where 1064 00:53:12,920 --> 00:53:16,000 we, I think, have a really good understanding of like how the 1065 00:53:16,000 --> 00:53:18,960 actual end user engineer or scientist runs all these work 1066 00:53:18,960 --> 00:53:20,560 flows. What are the problems they're 1067 00:53:20,560 --> 00:53:22,360 trying to solve? It can help them do that, you 1068 00:53:22,360 --> 00:53:24,560 know, 10 times better. Everybody wants, right? 1069 00:53:24,840 --> 00:53:28,280 And this entire compute conversation, right, becomes 1070 00:53:28,280 --> 00:53:31,920 pretty secondary, right, Because it's sort of like, you know, 1071 00:53:31,920 --> 00:53:34,280 like that's just it's just electricity, right? 1072 00:53:34,280 --> 00:53:37,480 It's it's like you're just powering the ability for an 1073 00:53:37,480 --> 00:53:39,560 engineer or scientist to come up with a cool innovation. 1074 00:53:40,640 --> 00:53:44,520 Well, that's kind of yeah, a good segue to talk about that a 1075 00:53:44,520 --> 00:53:46,560 little bit more. The you're right, we basically 1076 00:53:46,560 --> 00:53:50,352 focused on HPC. HPC was pre-AI, the big 1077 00:53:50,352 --> 00:53:54,292 changer, you know, you have more HPC run your simulation faster 1078 00:53:54,292 --> 00:53:56,268 and it still is, don't get me wrong. 1079 00:53:56,276 --> 00:54:02,192 And of course you know, GPUs help that, but the AI feels very 1080 00:54:02,192 --> 00:54:04,160 transformative for engineering. 1081 00:54:04,160 --> 00:54:09,200 So I mean, how have you seen that convergence on the Rescale 1082 00:54:09,200 --> 00:54:11,600 platform and from customers that you're speaking to, You know, 1083 00:54:11,600 --> 00:54:16,120 are they still just dabbling and with their sort of more in the 1084 00:54:16,120 --> 00:54:20,994 R&D phase, how much more sort of AI training or AI surrogates 1085 00:54:20,994 --> 00:54:24,320 are they doing next to their traditional CAE? 1086 00:54:24,320 --> 00:54:27,080 How are you seeing that you're, you're at the front with 1087 00:54:27,080 --> 00:54:28,920 customers? How is that transition? 1088 00:54:29,640 --> 00:54:34,680 The yeah, it's happening really fast, you know, are let's say 1089 00:54:34,680 --> 00:54:37,680 our most innovative and fast adopting customers. 1090 00:54:37,680 --> 00:54:41,960 They have AI based automation. You called agentic kind of 1091 00:54:42,440 --> 00:54:44,840 things built in. They have they're definitely 1092 00:54:44,840 --> 00:54:47,760 doing AI surrogates. They are really thinking ahead 1093 00:54:47,760 --> 00:54:51,280 of like what's the sort of let's just assume we already have AI 1094 00:54:51,280 --> 00:54:52,920 surrogates, right? Like what are the next level of 1095 00:54:52,920 --> 00:54:55,120 problems we can solve? Because once you have AI 1096 00:54:55,120 --> 00:54:57,880 surrogates and sort of, you know, domain I'm I'm pretty 1097 00:54:57,880 --> 00:55:00,557 familiar with is things like in the, in the multi physics 1098 00:55:00,557 --> 00:55:02,280 space, right? Like you're always trying to 1099 00:55:02,280 --> 00:55:04,320 kind of to reduced-order modelling or some way to 1100 00:55:04,320 --> 00:55:08,120 simplify these highly complex kind of large scale models? 1101 00:55:08,120 --> 00:55:10,320 And sort of. If you're say building an 1102 00:55:10,320 --> 00:55:12,960 airplane, you have this conceptual design phase where 1103 00:55:12,960 --> 00:55:16,440 you're like, it's smart to use like carbon fiber as a material 1104 00:55:16,440 --> 00:55:19,320 system for this airplane where all the implications. 1105 00:55:19,960 --> 00:55:22,800 The challenge with that is always like we get to the 1106 00:55:22,800 --> 00:55:26,960 detailed design that that's when all the details like it really 1107 00:55:26,960 --> 00:55:30,160 matter. And it'll be like, you know, at 1108 00:55:30,160 --> 00:55:32,360 the at the base of like a commercial aircraft, you might 1109 00:55:32,360 --> 00:55:34,560 have 100 to 200 plies of carbon fiber. 1110 00:55:34,560 --> 00:55:36,560 And then you have this like bolt that comes through there into 1111 00:55:36,560 --> 00:55:39,040 the titanium fitting. And there's like a bunch of 1112 00:55:39,040 --> 00:55:41,360 rules on like how that bolt supposed to interact with the 1113 00:55:41,360 --> 00:55:43,000 system and all this kind of stuff, right? 1114 00:55:43,960 --> 00:55:48,320 And it'll be like, OK, like we really need 200 carbon fiber. 1115 00:55:48,320 --> 00:55:51,880 Plies at that root, right? But then it'll be like, well, we 1116 00:55:51,880 --> 00:55:55,320 can't just go from 200 carbon fiber plies to like 10 in the 1117 00:55:55,320 --> 00:55:57,920 next panel over. Like this needs to go very sort 1118 00:55:57,920 --> 00:56:03,127 of slowly, like reduce it by 10% sort of every panel because 1119 00:56:03,127 --> 00:56:05,348 you can't have this disproportionate kind of 1120 00:56:05,348 --> 00:56:07,260 stress situation, right? And then it's like, oh, well, 1121 00:56:07,268 --> 00:56:09,432 this panel, we need actually be able to make it. 1122 00:56:09,440 --> 00:56:13,427 And you know, like our tool forming process can only handle 1123 00:56:13,427 --> 00:56:17,325 XYZ. So there's a lot more that 1124 00:56:17,333 --> 00:56:21,712 detail drives an enormous amount of the actual design. 1125 00:56:21,720 --> 00:56:25,420 And the problem has always been is like in that conceptual 1126 00:56:25,428 --> 00:56:28,168 design phase, how the historic has been done, you get a lot of 1127 00:56:28,168 --> 00:56:29,160 really smart people in the room. 1128 00:56:29,400 --> 00:56:31,920 And you say, yeah, this seems like the right decision because 1129 00:56:31,920 --> 00:56:34,160 I've seen some test data on carbon fiber over here and I've 1130 00:56:34,160 --> 00:56:35,680 like run some experiments over there. 1131 00:56:35,720 --> 00:56:39,280 But what you really want to do is kind of take all that really 1132 00:56:39,280 --> 00:56:43,560 detailed complex information and synthesize it up to help make 1133 00:56:43,600 --> 00:56:46,440 one of these kind of higher level decisions, right? 1134 00:56:46,480 --> 00:56:49,800 And what AI surrogates allow you to do is to simplify that 1135 00:56:49,800 --> 00:56:52,880 compression process. I do agree that it is, you know, 1136 00:56:52,880 --> 00:56:54,680 the criticism is like, hey, this is misleading. 1137 00:56:54,680 --> 00:56:57,040 You can, you can run the simulation 10,000 times faster, 1138 00:56:57,040 --> 00:56:58,920 etcetera. I mean, it is, it depends on 1139 00:56:58,920 --> 00:57:00,920 your frame of reference whether that's true or not. 1140 00:57:00,920 --> 00:57:05,650 But like what what is absolutely true is that using neural 1141 00:57:05,650 --> 00:57:09,614 networks to compress this highly complex information, right and 1142 00:57:09,614 --> 00:57:12,776 solve next sort of generation problems that you would never 1143 00:57:12,776 --> 00:57:15,320 even attempt before, right? Are now possible, right? 1144 00:57:15,320 --> 00:57:19,400 And and you have this much more elegant process to sort of 1145 00:57:19,400 --> 00:57:21,760 capture the intelligence of your engineering organization. 1146 00:57:22,680 --> 00:57:25,240 So and you can look do that at the, you know, individual 1147 00:57:25,240 --> 00:57:26,244 discipline solver level, right? 1148 00:57:26,252 --> 00:57:28,424 But you can also do that at higher levels. 1149 00:57:28,432 --> 00:57:30,460 And then you can start incorporating, you know, these 1150 00:57:30,468 --> 00:57:32,481 sort of things like manufacturing constraints and 1151 00:57:32,489 --> 00:57:35,200 things like that. And if you can, you know, the, 1152 00:57:35,208 --> 00:57:38,268 the faster you can do that loop basically, right, like the 1153 00:57:38,276 --> 00:57:41,170 better design you will get. Like I think what people are 1154 00:57:41,178 --> 00:57:43,740 usually surprised by is that that loop, there's exceptions to 1155 00:57:43,740 --> 00:57:46,594 this, but that loop takes a long time. 1156 00:57:46,602 --> 00:57:50,560 Like in in typical aerospace company, both Airbus and Boeing 1157 00:57:50,560 --> 00:57:54,420 of this sort of like loop of just like all the different 1158 00:57:54,428 --> 00:57:57,444 engineering disciplines to sort of say, OK, to have this shape 1159 00:57:57,452 --> 00:58:00,082 wing based on this ship, we're going to have these like loads 1160 00:58:00,090 --> 00:58:02,260 in the system based on these loads, like this sort of 1161 00:58:02,268 --> 00:58:04,826 mechanical team is going to like decide what kind of like 1162 00:58:04,834 --> 00:58:07,295 structure we can do because of the structure, it's going to 1163 00:58:07,303 --> 00:58:09,412 bend a certain way, which then comes back to like, OK, that's 1164 00:58:09,420 --> 00:58:10,870 the shape of the wing at cruise, right? 1165 00:58:10,878 --> 00:58:14,534 And so you have this big loop and that loop takes like three 1166 00:58:14,542 --> 00:58:16,978 to four months. It could take like an hour, 1167 00:58:16,986 --> 00:58:18,805 right? Like if you sort of said, hey, 1168 00:58:18,813 --> 00:58:22,165 I'm not constrained by compute. You're just running a whole 1169 00:58:22,173 --> 00:58:24,706 bunch of software, right? You still need some really smart 1170 00:58:24,706 --> 00:58:26,054 people to make some smart decisions, right? 1171 00:58:26,062 --> 00:58:28,420 Like it doesn't totally eliminate the engineer. 1172 00:58:28,428 --> 00:58:31,239 No. But, but I do think that that's 1173 00:58:31,239 --> 00:58:34,712 where the industry is going. Our customers like the the 1174 00:58:34,720 --> 00:58:37,704 bleeding edge customers, they are implementing all this stuff 1175 00:58:37,704 --> 00:58:40,075 right now and it's all possible today. 1176 00:58:40,083 --> 00:58:42,518 Like the exciting thing now is like, hey, it's right there. 1177 00:58:42,526 --> 00:58:46,060 You just have to do it. That's exactly that's the way I 1178 00:58:46,060 --> 00:58:48,923 normally describe it to people is that, you know, 10 years ago 1179 00:58:48,923 --> 00:58:52,922 there used to be this dream of real time CAE that, you know, 1180 00:58:52,930 --> 00:58:56,650 HPC would become so fast that you could do the simulation in 1181 00:58:56,658 --> 00:58:58,845 real time. But it actually is impossible. 1182 00:58:58,853 --> 00:59:01,620 You know, there's various reasons why a traditional 1183 00:59:01,628 --> 00:59:05,605 structures or fluid solver will never really become real time. 1184 00:59:05,613 --> 00:59:10,040 And that's where it's, I feel the AI surrogates, because by 1185 00:59:10,048 --> 00:59:14,520 definition they can be real time, does allow for the more 1186 00:59:14,528 --> 00:59:16,935 agentic AI. Because how can you really have 1187 00:59:16,935 --> 00:59:20,140 an agentic AI system when one of the agents is a solver that 1188 00:59:20,148 --> 00:59:24,023 takes 12 hours to run? Like, yeah, you can do that, but 1189 00:59:24,023 --> 00:59:26,903 it, it doesn't feel like the end goal. 1190 00:59:26,911 --> 00:59:30,620 The end goal should be like Google Gemini just because I 1191 00:59:30,628 --> 00:59:34,368 happen to use that one. There's others where like I want 1192 00:59:34,368 --> 00:59:37,316 to ask it to do something for me. 1193 00:59:37,324 --> 00:59:40,264 I don't want to wait 12 hours to get the answer back. 1194 00:59:40,272 --> 00:59:44,178 It feels like you want it to come in a short time and I feel 1195 00:59:44,178 --> 00:59:47,339 like the AI surrogates plug into that AI agent theme more 1196 00:59:47,339 --> 00:59:49,600 easily if you know what I mean. 1197 00:59:49,600 --> 00:59:53,000 I don't know if you agree. Yeah, it's a good. 1198 00:59:53,000 --> 00:59:57,144 It's an interesting way for any and I guess the way I look at 1199 00:59:57,144 --> 01:00:00,560 it is you have kind of, you know, just like you have deep 1200 01:00:00,560 --> 01:00:02,960 research, it takes a lot longer and then you have also like 1201 01:00:02,960 --> 01:00:04,840 quick responses. It's just like that, right? 1202 01:00:04,840 --> 01:00:08,947 Like where it's like, hey, if I am a designer and I just want 1203 01:00:08,947 --> 01:00:14,600 to change the shape of a mirror, right, I can get an instant 1204 01:00:14,600 --> 01:00:17,760 aero response that's like 99% accurate. 1205 01:00:18,400 --> 01:00:21,400 That is awesome, right? Like before, my alternative was 1206 01:00:21,480 --> 01:00:25,063 I changed the design, sent the CAD model over to the aero team. 1207 01:00:25,063 --> 01:00:27,120 Three days later they send me 1208 01:00:27,120 --> 01:00:29,720 back. Hey, that, that create a lot of 1209 01:00:29,720 --> 01:00:34,480 drag, right? So now you can get like a like a 1210 01:00:34,480 --> 01:00:37,760 pretty good answer and and pretty good. 1211 01:00:37,760 --> 01:00:41,520 Like I've seen cases where this is like over 99.9% accurate and 1212 01:00:41,520 --> 01:00:43,800 like, of course that depends on like the data you train on and 1213 01:00:43,800 --> 01:00:46,600 all these kind of things. But ultimately you can get these 1214 01:00:46,600 --> 01:00:50,600 responses back that dramatically reduce the cycle. 1215 01:00:50,600 --> 01:00:53,350 So you just eliminated 3 days by providing like a rough 1216 01:00:53,350 --> 01:00:54,752 estimate that's almost real time. 1217 01:00:54,760 --> 01:00:57,780 I still think that the detailed simulations do really matter. 1218 01:00:57,788 --> 01:00:59,880 Like that's your training data, right? 1219 01:00:59,888 --> 01:01:02,345 Like so like, and there's a couple of parts there, right? 1220 01:01:02,353 --> 01:01:05,842 Like I think there's the, I call it the sim-to-real gap, which is 1221 01:01:05,842 --> 01:01:08,685 a separate discussion. But that's, it's another, I 1222 01:01:08,693 --> 01:01:11,688 think, interesting element, which is like, OK, are these 1223 01:01:11,696 --> 01:01:14,300 simulations that we're running, how accurate are they? 1224 01:01:14,308 --> 01:01:15,852 Because the real ground-truth data, right? 1225 01:01:15,860 --> 01:01:17,794 It's not even wind tunnel testing. 1226 01:01:17,802 --> 01:01:19,716 It's actually like the real data, right? 1227 01:01:19,724 --> 01:01:21,390 And the wind tunnel is a proxy for that. 1228 01:01:21,398 --> 01:01:24,230 And then your simulation tool is a proxy for the wind tunnel. 1229 01:01:24,238 --> 01:01:26,352 And yes, all the physics equations are correct, but 1230 01:01:26,360 --> 01:01:29,068 there's like a lot of little things that still can, can 1231 01:01:29,076 --> 01:01:30,623 change, like real world outcomes. 1232 01:01:30,631 --> 01:01:34,960 And so, but from the pure simulation, you now have the 1233 01:01:34,968 --> 01:01:38,068 ability to generate and ground all this data like very 1234 01:01:38,076 --> 01:01:39,740 efficiently with these AI models, right? 1235 01:01:39,748 --> 01:01:43,460 And so like, yes, that designer doesn't want to wait for like an 1236 01:01:43,460 --> 01:01:46,750 like, yeah, you can give them an agentic capability spin off a 1237 01:01:46,750 --> 01:01:49,463 simulation, right? That's not that useful to them, 1238 01:01:49,463 --> 01:01:53,139 right? Where agentic to me means, and 1239 01:01:53,147 --> 01:01:55,968 you know, it's a kind of a buzzword. 1240 01:01:55,976 --> 01:01:57,436 So everything's agentic these days. 1241 01:01:57,444 --> 01:02:01,327 But to me it's more of this kind of like like proactive thing, 1242 01:02:01,335 --> 01:02:03,227 right? So where it's like, OK, here, 1243 01:02:03,235 --> 01:02:07,308 here's what happened. And it'll sort of describe a 1244 01:02:07,316 --> 01:02:09,400 customer scenario, automotive OEM, right? 1245 01:02:09,408 --> 01:02:14,250 And they have a supplier, supplier say it's the seat 1246 01:02:14,258 --> 01:02:18,294 manufacturer, they changed the design of the seat because some 1247 01:02:18,294 --> 01:02:20,472 reason for same manufacturability on their end, 1248 01:02:20,472 --> 01:02:23,708 right? That design change is is sort of 1249 01:02:23,708 --> 01:02:27,900 automatically synced this OEM, this is happening because this 1250 01:02:27,908 --> 01:02:31,196 seat manufacturer is in Europe and the OEM is in the US. 1251 01:02:31,204 --> 01:02:34,402 You know, this is happening at like 2:00 AM right now because 1252 01:02:34,410 --> 01:02:37,858 that change happened. You need to rerun your crash 1253 01:02:37,866 --> 01:02:39,194 analysis. So you could do an instant 1254 01:02:39,202 --> 01:02:41,864 response of like sort of an AI surrogate of like, hey, is this 1255 01:02:41,864 --> 01:02:44,500 going to be a is this look great or is this bad? 1256 01:02:44,508 --> 01:02:46,005 And it's like this might be a problem. 1257 01:02:46,013 --> 01:02:49,116 OK, now because of that and the simulation engineer sleeping 1258 01:02:49,124 --> 01:02:54,008 because of that, a spin up like LS-DYNA job to go run the crash 1259 01:02:54,008 --> 01:02:57,530 analysis and I'm just going to run it for just this component, 1260 01:02:57,530 --> 01:03:00,000 the subcomponents, you know, in a way that's like pretty 1261 01:03:00,008 --> 01:03:01,166 efficient based on a bunch of rules. 1262 01:03:01,174 --> 01:03:04,020 I gave that agent right, which said, hey, you can't spend more 1263 01:03:04,020 --> 01:03:06,680 than $1000 and if it design change like this comes in, 1264 01:03:06,688 --> 01:03:08,098 always use the AI surrogate first. 1265 01:03:08,106 --> 01:03:12,460 And now I, I want to, you know, like failed that test, though, 1266 01:03:12,468 --> 01:03:13,980 right then then run this simulation. 1267 01:03:13,988 --> 01:03:17,220 So when I come in in the morning, right open my laptop at 1268 01:03:17,220 --> 01:03:20,460 7 a.m., I've got right there. I have the simulation results 1269 01:03:20,468 --> 01:03:22,880 right to go review right? And it's like it's. 1270 01:03:22,888 --> 01:03:25,471 Pre done the preprocessing, this whole like trace of all these 1271 01:03:25,471 --> 01:03:27,605 things that happened is presented to me in like a simple 1272 01:03:27,605 --> 01:03:30,952 way. And you know, my job just got so 1273 01:03:30,952 --> 01:03:34,244 much faster and more efficient, right, Like, like the 1274 01:03:34,244 --> 01:03:37,250 real way this happened before and and most organizations work 1275 01:03:37,250 --> 01:03:40,008 this way is like, oh, this this CAD file got zipped up, 1276 01:03:40,008 --> 01:03:42,320 right, and then it's like sat in somebody's outbox, but they 1277 01:03:42,320 --> 01:03:44,464 were in Europe, so they were actually on vacation for another 1278 01:03:44,464 --> 01:03:46,695 week, right. And so like it didn't actually 1279 01:03:46,703 --> 01:03:50,152 get make it over to the OEM because like whatever random 1280 01:03:50,160 --> 01:03:53,960 human delay, right, then the zip file comes over and then 1281 01:03:53,968 --> 01:03:55,570 somebody has to do this analysis, right? 1282 01:03:55,578 --> 01:03:58,960 And then, and then all these sort of steps happen, right? 1283 01:03:58,968 --> 01:04:02,636 And I think if you can sort of shorten that entire process now, 1284 01:04:02,636 --> 01:04:05,778 it's not perfect, right? Like if you look at the tools we 1285 01:04:05,778 --> 01:04:08,400 have today, there's a lot of expertise to make all these 1286 01:04:08,408 --> 01:04:10,029 judgements and do all these things. 1287 01:04:10,037 --> 01:04:13,255 But you can already see today that like our customers are 1288 01:04:13,263 --> 01:04:16,010 already doing this. They're automating more and more 1289 01:04:16,010 --> 01:04:19,095 of this process, right? And you're just going to get 1290 01:04:19,103 --> 01:04:20,880 comfortable with it. It's just like you get 1291 01:04:20,888 --> 01:04:23,655 comfortable with using a Gemini or ChatGPT or whatever, right? 1292 01:04:23,663 --> 01:04:26,740 And that's the sort of agentic thing is sort of like. 1293 01:04:26,748 --> 01:04:28,500 Proactively doing all the kind of like. 1294 01:04:28,508 --> 01:04:31,188 Grunt work that you don't want to do right? 1295 01:04:31,196 --> 01:04:35,264 Yeah, but I, I seriously, I don't know if I'm just, you 1296 01:04:35,272 --> 01:04:38,866 know, drinking the Kool-aid or whatever the phrase is, but I, I 1297 01:04:38,866 --> 01:04:43,016 really can't see how this will just totally change because. 1298 01:04:43,024 --> 01:04:48,172 Well, the first thing to say maybe what, what I was 1299 01:04:48,180 --> 01:04:51,980 mentioning before is even if I believe that you, the AI 1300 01:04:51,988 --> 01:04:56,580 surrogates, make it good enough, you're still going to have a 1301 01:04:56,588 --> 01:05:00,040 whole load of huge HPC resources to create training data. 1302 01:05:00,048 --> 01:05:02,840 You know, so even if the surrogate models become good 1303 01:05:02,848 --> 01:05:05,370 enough, and unless there's some breakthrough that I'm not aware 1304 01:05:05,370 --> 01:05:07,840 of, you still need all the traditional stuff to, to 1305 01:05:07,848 --> 01:05:11,062 generate the data. So actually it's, it's not a, 1306 01:05:11,070 --> 01:05:15,382 it's a replacement in that the end user may primarily use it, 1307 01:05:15,390 --> 01:05:18,239 but in the back end, somebody's still having to generate this 1308 01:05:18,247 --> 01:05:21,120 data to go and train the models so that you're not truly 1309 01:05:21,128 --> 01:05:22,642 replacing it. It may just be in the 1310 01:05:22,650 --> 01:05:26,142 background, if you know what I mean, but I really feel that 1311 01:05:26,150 --> 01:05:30,882 the, the AI engineer thing seems more and more real because what 1312 01:05:30,882 --> 01:05:34,612 you just described to me is eventually through enough 1313 01:05:34,620 --> 01:05:38,360 training or awareness of different agents, isn't it just 1314 01:05:38,360 --> 01:05:43,650 that I become like a manager? So I have I and I have a team of 1315 01:05:43,650 --> 01:05:47,272 people who are not real people who are doing this analysis for 1316 01:05:47,272 --> 01:05:49,820 me and I'm just looking at it at the end. 1317 01:05:49,828 --> 01:05:51,275 Yes, and I think it's awesome, right? 1318 01:05:51,283 --> 01:05:54,830 Like I think it's so so the other part's like this is 1319 01:05:54,838 --> 01:05:57,900 already happening today. So hopefully shortly we'll be 1320 01:05:57,908 --> 01:06:00,506 able to share some like maybe public customer case studies and 1321 01:06:00,506 --> 01:06:03,204 exactly how this works and have the customers talk about 1322 01:06:03,204 --> 01:06:05,520 it. But I think, you know, there's 1323 01:06:05,520 --> 01:06:07,600 no part what I just described, right? 1324 01:06:07,600 --> 01:06:10,572 Like all all the components that we have for any company to go 1325 01:06:10,572 --> 01:06:13,760 do that, you don't need a magic AI surrogate model, right? 1326 01:06:13,800 --> 01:06:20,000 Like it's just like, but I think the right framing is more AI 1327 01:06:20,000 --> 01:06:23,760 surrogates is 1 module of AI. There's many AI modules in the 1328 01:06:23,760 --> 01:06:26,120 entire product development engineering process, right? 1329 01:06:26,960 --> 01:06:29,600 Some of these are more leveraging LLMs, right? 1330 01:06:29,600 --> 01:06:32,000 Like some of them are more like this compute recommendation 1331 01:06:32,000 --> 01:06:33,440 engine. That's like a totally different 1332 01:06:33,440 --> 01:06:35,200 category of AI. But it helps solve this like 1333 01:06:35,200 --> 01:06:36,720 search problem much more efficiently. 1334 01:06:37,760 --> 01:06:40,640 You have AI surrogates who do these sort of probabilistic 1335 01:06:40,640 --> 01:06:42,440 physics analysis much more efficiently. 1336 01:06:42,440 --> 01:06:46,950 The, the really nice part is the productization of these 1337 01:06:46,950 --> 01:06:49,524 because like if I tried to steelman, sort of the cynics 1338 01:06:49,524 --> 01:06:52,690 and I've I've heard many of them in sort of the CAE space, 1339 01:06:52,690 --> 01:06:55,569 right where it's like, but those are not real like engineering 1340 01:06:55,569 --> 01:06:57,388 calculations. These are just like 1341 01:06:57,396 --> 01:06:59,150 approximations, etcetera. Approximations been around 1342 01:06:59,158 --> 01:07:00,520 forever. We've always done surrogate. 1343 01:07:00,528 --> 01:07:04,260 Modeling right like the. I think what's really important 1344 01:07:04,260 --> 01:07:06,040 is like, solve the real business problem, right? 1345 01:07:06,048 --> 01:07:08,928 Like like solve actually what, why are people even doing this 1346 01:07:08,936 --> 01:07:10,586 kind of work, right? Like, well, they're just trying 1347 01:07:10,586 --> 01:07:12,672 to figure out the sort of physics answer for something. 1348 01:07:12,680 --> 01:07:16,462 And almost all simulations that are run are essentially a waste 1349 01:07:16,462 --> 01:07:19,864 because in the end, there's only like one or two that really 1350 01:07:19,864 --> 01:07:22,765 matter that go into certification and like the aero 1351 01:07:22,765 --> 01:07:25,725 model for say a Boeing airplane, right? 1352 01:07:25,733 --> 01:07:29,456 Like they're sort of if you want the entire way of how did you 1353 01:07:29,464 --> 01:07:31,334 get there, right? Like, yes, then all those other 1354 01:07:31,334 --> 01:07:33,580 data matters, right? Like all the other experiments 1355 01:07:33,588 --> 01:07:37,495 that were run, but more than like 99% of all those 1356 01:07:37,503 --> 01:07:40,024 computations that don't impact this, like final design is 1357 01:07:40,032 --> 01:07:42,160 great, right? And so the question is just 1358 01:07:42,168 --> 01:07:44,080 like, how can you go through that search process as quickly 1359 01:07:44,088 --> 01:07:46,490 and as efficiently as possible? And what changes everything is, 1360 01:07:46,490 --> 01:07:49,780 I think it's if you can solve that search problem so much 1361 01:07:49,788 --> 01:07:52,974 faster, it changes how you do engineering completely, right? 1362 01:07:52,982 --> 01:07:55,360 And not in a negative way, right? 1363 01:07:55,368 --> 01:07:58,270 Like to me because, because there's also like the other 1364 01:07:58,278 --> 01:08:00,800 criticism on this is like, OK, well then we just replace all 1365 01:08:00,808 --> 01:08:02,272 these engineers. Like I'm an engineer, like why 1366 01:08:02,280 --> 01:08:03,545 are you replacing me? Right. 1367 01:08:03,553 --> 01:08:09,386 But this is all like, what is your frame on technology and 1368 01:08:09,394 --> 01:08:13,372 like how the world should kind of work, etcetera, right? 1369 01:08:13,380 --> 01:08:18,935 But like it's I think right it it's super important to kind of 1370 01:08:18,935 --> 01:08:22,452 like jump on these waves as like this is awesome. 1371 01:08:22,460 --> 01:08:25,642 Why is it awesome? Well, what we just have to grind 1372 01:08:25,642 --> 01:08:28,899 through a whole bunch of CAD and CAE modeling with like a 1373 01:08:28,899 --> 01:08:32,268 team of 12 people for like 6 months to get to this answer, 1374 01:08:32,268 --> 01:08:35,895 not to get the exact same answer with two people in a 1375 01:08:35,895 --> 01:08:38,920 week, right? That is a win for everybody. 1376 01:08:38,960 --> 01:08:42,279 Those ten other people, right, that you may not need any more 1377 01:08:42,279 --> 01:08:45,680 can go do five other projects. Those ten other people can do 1378 01:08:46,240 --> 01:08:48,600 more meaningful things, learn other things, right? 1379 01:08:49,160 --> 01:08:52,479 If you look at the history of like CAD modelling, right? 1380 01:08:52,640 --> 01:08:55,000 Like it used to be, we have these drafters I'm sure you've 1381 01:08:55,000 --> 01:08:59,000 seen these pictures of like, you know, yeah, 100 people just 1382 01:08:59,319 --> 01:09:02,560 drafting the physical documents, right, To be able to build 1383 01:09:02,560 --> 01:09:06,040 components. I don't think it's a negative 1384 01:09:06,040 --> 01:09:07,960 thing that people don't need to do this anymore. 1385 01:09:08,000 --> 01:09:11,520 Like you press a button in a 3D CAD product and spits out like 1386 01:09:11,520 --> 01:09:13,920 so. It's a productivity argument, 1387 01:09:13,920 --> 01:09:16,520 isn't it? You know, yeah, you can do more 1388 01:09:16,520 --> 01:09:18,359 efficient. I mean, there's a debate on 1389 01:09:18,359 --> 01:09:20,359 jobs. I guess it's maybe the correct 1390 01:09:20,359 --> 01:09:22,200 thing to say is it's not the job. 1391 01:09:22,439 --> 01:09:25,279 There will be people's jobs that they don't need to do anymore, 1392 01:09:25,279 --> 01:09:28,160 but you'd hope they could trans transition to a different job. 1393 01:09:28,279 --> 01:09:30,399 I guess is. Yes, and like it's. 1394 01:09:30,800 --> 01:09:34,191 You know, the U.S. I think the job retraining 1395 01:09:34,191 --> 01:09:38,640 stat is something like 20 to 30% do reskilling every year. 1396 01:09:38,640 --> 01:09:41,520 Yeah, yeah. So yes, unemployment stays at 1397 01:09:41,520 --> 01:09:44,680 like X level, right. But like in the US statistic. 1398 01:09:44,680 --> 01:09:46,080 But I imagine it's similar globally. 1399 01:09:48,439 --> 01:09:53,359 And I don't personally find it very satisfying if you're like 1400 01:09:53,359 --> 01:09:57,440 if your job was like, say, just drafting documents, right? 1401 01:09:57,440 --> 01:09:59,760 But actually, like, you can just automate it with software. 1402 01:10:01,040 --> 01:10:02,240 It's right. Not a very fun job. 1403 01:10:02,440 --> 01:10:04,600 Right. Like no, no, no, hadn't. 1404 01:10:04,720 --> 01:10:08,240 There's just certain stuff. I think that machines or AI now 1405 01:10:08,240 --> 01:10:10,480 can do better. You know, if I have to manually 1406 01:10:10,880 --> 01:10:15,920 go through 50 simulations and try and write a report, why this 1407 01:10:15,920 --> 01:10:20,400 one moved the vortex here and AI could do exactly the same task 1408 01:10:20,760 --> 01:10:23,920 in less than a minute. And then you can still write the 1409 01:10:23,920 --> 01:10:25,400 report and analyze and think about it. 1410 01:10:25,400 --> 01:10:27,920 But you didn't have to do some of that manual stuff that 1411 01:10:27,920 --> 01:10:31,280 probably took you half a day. To do exactly so we've we've 1412 01:10:31,280 --> 01:10:32,360 actually done this with our customers. 1413 01:10:32,360 --> 01:10:34,280 We survey them, right. Let's say what part about your 1414 01:10:34,280 --> 01:10:37,040 job do you dislike do or where you don't feel like you're 1415 01:10:37,040 --> 01:10:38,560 adding value? One of the top things that come 1416 01:10:38,560 --> 01:10:40,560 up is exactly what you just described is write all these 1417 01:10:40,560 --> 01:10:43,760 reports. Guess what, it's really easy to 1418 01:10:43,760 --> 01:10:46,240 take a bunch of like simulation files, right? 1419 01:10:46,240 --> 01:10:48,367 You've decided, hey, this is the right sort of answer and 1420 01:10:48,367 --> 01:10:49,960 here's sort of the high level reasons why. 1421 01:10:51,040 --> 01:10:54,560 And you can just generate a report using AI tool that works 1422 01:10:54,560 --> 01:10:56,680 today. There's no reason not to do 1423 01:10:56,680 --> 01:10:57,640 that. Now, do you need to review the 1424 01:10:57,640 --> 01:11:00,360 report? Yes, right, but like you can 1425 01:11:00,360 --> 01:11:05,160 shortcut a lot of the process, which actually allows the person 1426 01:11:05,240 --> 01:11:08,040 or the human to add, you know, spend much more time on those 1427 01:11:08,040 --> 01:11:10,680 like value added tasks, right. And and I kind of think of this 1428 01:11:10,680 --> 01:11:14,668 always as like moving up a layer of abstraction where now you 1429 01:11:14,668 --> 01:11:16,200 can do so much more. I think what people 1430 01:11:16,200 --> 01:11:19,120 underestimate is like when you change something, right, like 1431 01:11:19,120 --> 01:11:23,400 say the six month process down to a week, that changes 1432 01:11:23,440 --> 01:11:25,360 everything, right? It changes like industry 1433 01:11:25,360 --> 01:11:27,640 dynamics because like you can now do product development so 1434 01:11:27,640 --> 01:11:30,280 much faster. It changes like what kind of 1435 01:11:30,320 --> 01:11:32,120 innovations you can actually build, right? 1436 01:11:32,840 --> 01:11:34,720 It changes like the org structures of companies, 1437 01:11:34,720 --> 01:11:37,800 obviously, but I think that's all, you know, like value added 1438 01:11:37,800 --> 01:11:41,000 to society, right? Like I think it's great, yeah. 1439 01:11:41,000 --> 01:11:43,560 And. I guess full circle. 1440 01:11:43,560 --> 01:11:47,451 The underpinning of all of this is access to compute, 1441 01:11:47,451 --> 01:11:48,868 basically. That's right, isn't it? 1442 01:11:48,876 --> 01:11:53,580 I do think, I mean, I, I will say like, you know, pre AI wave, 1443 01:11:53,580 --> 01:11:55,460 right? And, and, and I think, yeah, 1444 01:11:55,468 --> 01:11:59,240 that's when you were probably at AWS, you know, compute was 1445 01:11:59,248 --> 01:12:02,540 really considered like very much a commodity that's like, yeah, 1446 01:12:02,548 --> 01:12:04,880 like there's all these like complexities, even HPC, but like 1447 01:12:04,880 --> 01:12:07,934 in the end, it's just like sort of a means to the end, 1448 01:12:07,934 --> 01:12:09,760 etcetera. I think AI has shown, which is 1449 01:12:09,760 --> 01:12:12,400 quite exciting how important compute still is, right? 1450 01:12:13,360 --> 01:12:15,400 And they're sort of the, the high level thing of like, hey, 1451 01:12:15,400 --> 01:12:17,760 well, you need this massive scale compute to solve this, 1452 01:12:17,760 --> 01:12:19,720 specifically this sort of LLM training model. 1453 01:12:19,720 --> 01:12:21,680 But I think that analogy applies to many other things. 1454 01:12:21,680 --> 01:12:24,120 So if you take a typical engineering problem at a large 1455 01:12:24,120 --> 01:12:27,280 company, a complex problem like like like high end aerospace 1456 01:12:27,280 --> 01:12:31,000 product, these people doing this work are highly compute bound, 1457 01:12:32,400 --> 01:12:34,640 right? And they're compute bound at 1458 01:12:34,640 --> 01:12:38,080 the, you know, I can only run so many simulations, but they're 1459 01:12:38,080 --> 01:12:40,160 also compute bound at like as you move up these layers of 1460 01:12:40,160 --> 01:12:42,280 abstraction, right? As I sort of like this agentic 1461 01:12:42,280 --> 01:12:46,520 engineering, it becomes easier and easier to actually like sort 1462 01:12:46,520 --> 01:12:47,940 of kick off a simulation, right? 1463 01:12:47,948 --> 01:12:50,575 Add more training data, all these kind of things, right? 1464 01:12:50,583 --> 01:12:53,358 That's where you need to be really smart about like, well, 1465 01:12:53,366 --> 01:12:55,469 what are all the, like design points you want to run? 1466 01:12:55,477 --> 01:12:57,721 Again, AI can help you do all these things, right? 1467 01:12:57,729 --> 01:13:02,392 But the compute bill sort of going up and up and up. 1468 01:13:02,400 --> 01:13:05,716 However, in most organizations, if you look at like, what is 1469 01:13:05,724 --> 01:13:08,220 your R&D spent, the biggest cost is the people. 1470 01:13:08,228 --> 01:13:10,455 Yeah. So this opportunity to get much 1471 01:13:10,455 --> 01:13:12,326 more leverage out of an individual, right? 1472 01:13:12,334 --> 01:13:16,692 Like they can just add much more value in this cycle is is 1473 01:13:16,700 --> 01:13:18,795 massive. I think it is underpinned by 1474 01:13:18,803 --> 01:13:21,642 compute. And if you don't have that 1475 01:13:21,650 --> 01:13:23,708 foundational building block, right, like you're, you're 1476 01:13:23,716 --> 01:13:25,456 highly constrained. The good thing is engineers can 1477 01:13:25,456 --> 01:13:26,335 always work around those constraints. 1478 01:13:26,343 --> 01:13:28,775 You know, they just make do with whatever you're given. 1479 01:13:28,783 --> 01:13:32,180 But you know, like those constraints, like really 1480 01:13:32,188 --> 01:13:33,834 constrained also innovation, right? 1481 01:13:33,842 --> 01:13:37,471 They constrain the ability, like if you can make it, SpaceX is a 1482 01:13:37,471 --> 01:13:39,556 good example, right? Like it like sort of back when I 1483 01:13:39,556 --> 01:13:41,920 was working at Boeing, it was like, hey, there's a SpaceX 1484 01:13:41,928 --> 01:13:44,608 company and, you know, Boeing thought they were pretty good at 1485 01:13:44,608 --> 01:13:47,088 like launching rockets and putting satellites in orbit. 1486 01:13:47,096 --> 01:13:50,624 And you know, at SpaceX there will take the cost. 1487 01:13:50,632 --> 01:13:54,860 I think of like putting a kilo in space, like down by like 20 1488 01:13:54,868 --> 01:13:59,284 to 50 X depending on like kind of what metrics you use, but a 1489 01:13:59,292 --> 01:14:02,240 minimum 20X, right? Once you do that, right? 1490 01:14:02,248 --> 01:14:04,126 Like it opens up this entire market, right? 1491 01:14:04,134 --> 01:14:07,318 Like it's like, OK, now you can do Starlink, now you can do all 1492 01:14:07,318 --> 01:14:09,347 kinds of other stuff in space, right? 1493 01:14:09,355 --> 01:14:11,880 And you want those expansion opportunities. 1494 01:14:11,888 --> 01:14:15,104 I do feel like a lot of industries are pretty stuck, 1495 01:14:15,112 --> 01:14:19,384 like the way Boeing builds an airplane or Airbus for that 1496 01:14:19,392 --> 01:14:21,344 matter, right? Is pretty stuck in sort of these 1497 01:14:21,344 --> 01:14:23,328 old ways, right? And I think of the way you can 1498 01:14:23,336 --> 01:14:25,782 get out of that is using all these new tools where it's sort 1499 01:14:25,782 --> 01:14:27,850 of like something is at least 10X better. 1500 01:14:27,858 --> 01:14:31,545 It totally changes not only like how fast you can do the product 1501 01:14:31,545 --> 01:14:33,308 development, but also like the markets you can serve. 1502 01:14:33,316 --> 01:14:35,372 It changes the way you work with your customers. 1503 01:14:35,380 --> 01:14:37,716 If you're say, an aircraft manufacturer right, and that 1504 01:14:37,724 --> 01:14:39,886 pushes society forward, I think it's great. 1505 01:14:39,894 --> 01:14:43,370 I, I don't know about you, but the older I get and the more I 1506 01:14:43,378 --> 01:14:46,212 work with enterprise and the more I work with startups I 1507 01:14:46,220 --> 01:14:49,680 have, and it's probably not practically possible, but I 1508 01:14:49,688 --> 01:14:54,800 often think, OK, if company A is a legacy enterprise company and 1509 01:14:54,800 --> 01:14:58,460 they're trying to innovate and come up with some new product. 1510 01:14:58,468 --> 01:15:02,415 I really do think their best suggestion is they should go and 1511 01:15:02,415 --> 01:15:06,226 spin off a startup, right? Because I sometimes feel as if 1512 01:15:06,234 --> 01:15:09,522 they cannot change into they should literally go and take 50 1513 01:15:09,522 --> 01:15:13,022 super smart people, fund it and let them do whatever the hell 1514 01:15:13,030 --> 01:15:14,909 they want and then integrate it back. 1515 01:15:14,917 --> 01:15:18,304 Do you know what I mean? Some companies are just too 1516 01:15:18,312 --> 01:15:20,827 stiff to really innovate. Yes. 1517 01:15:20,835 --> 01:15:21,385 I don't. Disagree. 1518 01:15:21,393 --> 01:15:25,304 I mean, I think it is much harder to be running like one of 1519 01:15:25,304 --> 01:15:28,260 those companies and like transform as an organization and 1520 01:15:28,260 --> 01:15:30,180 it's proven through all the numbers, right? 1521 01:15:30,188 --> 01:15:33,640 Like like sort of the Fortune 500 companies, like, you know, 1522 01:15:33,648 --> 01:15:37,060 the stock market top companies rotate like very quickly, right? 1523 01:15:37,060 --> 01:15:38,400 Like most companies don't last 1524 01:15:38,400 --> 01:15:40,920 like 15–20 years as a public market company. 1525 01:15:41,280 --> 01:15:46,265 So I will say though, like, you know, like we as Boom 1526 01:15:46,265 --> 01:15:48,540 Supersonic is one of our customers, right, man, 1527 01:15:48,540 --> 01:15:50,126 starting a supersonic jet company, right? 1528 01:15:50,134 --> 01:15:53,832 I always think Rescale is like a pretty tough company to run, but 1529 01:15:53,832 --> 01:15:56,410 that is hard, right? But that really takes a lot of 1530 01:15:56,418 --> 01:15:57,820 courage, right, to do something like that. 1531 01:15:57,828 --> 01:16:00,400 So it's very admirable for folks who have the mission and willing 1532 01:16:00,400 --> 01:16:04,020 to kind of take that on. But what is also true is that 1533 01:16:04,028 --> 01:16:05,656 like, so they're super innovative. 1534 01:16:05,664 --> 01:16:06,903 They're adopting all these methods. 1535 01:16:06,911 --> 01:16:09,776 They're going to be way faster than any other kind of large 1536 01:16:09,784 --> 01:16:12,525 aerospace manufacturer trying to do a supersonic jet, of course, 1537 01:16:12,525 --> 01:16:14,327 but it's still really hard, right? 1538 01:16:14,335 --> 01:16:16,848 Like, like startups are just really hard, right? 1539 01:16:16,856 --> 01:16:22,080 And, you know, I'm, I would say I'm glad I'm, I work in the like 1540 01:16:22,080 --> 01:16:24,304 field of software, right? But we, our customers are all 1541 01:16:24,312 --> 01:16:27,995 building hardware pretty much. So it's, yeah, it's awesome to 1542 01:16:28,003 --> 01:16:31,948 be able to serve them. But I think it's, well, that's 1543 01:16:31,956 --> 01:16:34,400 the right answer. It is still like, you know, the, 1544 01:16:34,400 --> 01:16:36,900 the the odds are sort of stacked against you when you 1545 01:16:36,900 --> 01:16:40,420 start a company, right? And so even today, like boom 1546 01:16:40,428 --> 01:16:42,658 supersonic, right? Like they, they, they flew that 1547 01:16:42,658 --> 01:16:44,680 airplane. It's, it's like, built with 1548 01:16:44,680 --> 01:16:46,605 Rescale. Awesome, right to test airplane, 1549 01:16:46,605 --> 01:16:49,080 they got to, they build the real one. 1550 01:16:49,088 --> 01:16:52,638 They got to get all these customers, you know, and it's, 1551 01:16:52,646 --> 01:16:57,528 it's, it's, it's a tough, yeah, it's a tough problem to solve as 1552 01:16:57,528 --> 01:17:01,756 a as a small company, right? So, but, but that is the way 1553 01:17:01,764 --> 01:17:04,228 like I do think cultural change, like culture really matters. 1554 01:17:04,236 --> 01:17:08,605 So one thing we focus on at Rescale is like not only making 1555 01:17:08,605 --> 01:17:11,468 sure our customers are adopting all these capabilities, right, 1556 01:17:11,476 --> 01:17:14,720 but like if you, if you work at Rescale, we have this. 1557 01:17:14,728 --> 01:17:18,920 So being an AI first company is completely changes how you run 1558 01:17:18,928 --> 01:17:22,189 the company yourself, right? And so we're very focused on 1559 01:17:22,197 --> 01:17:23,438 that. Like this whole thing you were 1560 01:17:23,446 --> 01:17:25,930 talking about, hey, now I'm managing a bunch of agents. 1561 01:17:25,938 --> 01:17:29,790 So every employee at Rescale is empowered to kind of like, you 1562 01:17:29,798 --> 01:17:33,872 know, run their agents, right? And, and we are using every 1563 01:17:33,880 --> 01:17:36,760 single AI framework, right? Like I'm testing all of them 1564 01:17:36,768 --> 01:17:41,576 usually in parallel at the same time, very rapidly changing 1565 01:17:41,584 --> 01:17:43,952 landscape. But I think the way you build a 1566 01:17:43,952 --> 01:17:46,909 company and the way you scale a company is quite different in 1567 01:17:46,917 --> 01:17:50,071 this like world of when you have this AI tooling. 1568 01:17:50,079 --> 01:17:53,742 And I do think as a as if you consider Rescale a big or small 1569 01:17:53,742 --> 01:17:56,864 company, but as a 200 person company versus say a 2000 person 1570 01:17:56,864 --> 01:17:58,869 company, we have a distinct advantage. 1571 01:17:58,877 --> 01:18:02,724 Our ability to adopt new tooling for this and sort of like dog 1572 01:18:02,732 --> 01:18:06,682 food and use all these tools ourselves is, you know, if we 1573 01:18:06,690 --> 01:18:10,420 sort of execute well, is it really fast, right? 1574 01:18:10,428 --> 01:18:12,734 Like we can, we can adopt new tools really fast. 1575 01:18:12,742 --> 01:18:16,842 That allows us to understand how our customers will also need to 1576 01:18:16,842 --> 01:18:19,688 like sort of change and adopt tooling like this, right? 1577 01:18:19,696 --> 01:18:21,160 And, and it's a very different business. 1578 01:18:21,168 --> 01:18:22,180 Our business is building software. 1579 01:18:22,188 --> 01:18:24,680 Their business might be building a vehicle, but a lot of the 1580 01:18:24,688 --> 01:18:27,015 principles are the same, right? Like it, it's sort of like you 1581 01:18:27,023 --> 01:18:29,632 have to really rethink how you do business and how you do 1582 01:18:29,640 --> 01:18:30,945 engineering and all these things. 1583 01:18:30,953 --> 01:18:33,560 But I think that's also why it's exciting, right? 1584 01:18:33,568 --> 01:18:35,794 Because it's like it is a new paradigm. 1585 01:18:35,802 --> 01:18:39,050 These shifts don't happen very often, right? 1586 01:18:39,058 --> 01:18:42,498 Like I don't know, if you asked me a decade ago, are we going to 1587 01:18:42,498 --> 01:18:45,000 pass the Turing test, I'd have been like, I don't think so, 1588 01:18:45,008 --> 01:18:47,204 right? I would have been very wrong. 1589 01:18:47,212 --> 01:18:51,312 But like I think just like the cloud shift like opened up these 1590 01:18:51,312 --> 01:18:53,760 like massive, massive markets and and massive new 1591 01:18:53,768 --> 01:18:56,144 opportunities, right And many businesses could not exist 1592 01:18:56,152 --> 01:18:58,625 without sort of concept of cloud etcetera. 1593 01:18:58,633 --> 01:19:02,315 AI, I do think is even bigger, right? 1594 01:19:02,323 --> 01:19:05,574 It is sort of like the probably the biggest one of our 1595 01:19:05,582 --> 01:19:07,370 generation. And that's awesome because like, 1596 01:19:07,370 --> 01:19:10,436 you know, I don't know if you think about this, but I'm like, 1597 01:19:10,436 --> 01:19:13,356 man, what was it like when like the I mean, I was a kid, like 1598 01:19:13,364 --> 01:19:16,424 the Internet sort of first came out and then you could all of a 1599 01:19:16,424 --> 01:19:19,128 sudden like do this, like, you know, you could sort of message 1600 01:19:19,128 --> 01:19:21,474 people across the world. It's like it seemed crazy, 1601 01:19:21,482 --> 01:19:24,738 right? And it was really a special 1602 01:19:24,746 --> 01:19:27,695 time. And I think those, you know, you 1603 01:19:27,695 --> 01:19:29,225 go through the high of these waves. 1604 01:19:29,233 --> 01:19:31,960 I think right now is a really special time, right where we're 1605 01:19:31,960 --> 01:19:34,980 like the bleeding edge of applying AI to engineering and 1606 01:19:34,988 --> 01:19:37,540 like how you build companies, everything's changing and 1607 01:19:37,548 --> 01:19:39,635 that's, I think that's really fun. 1608 01:19:39,643 --> 01:19:44,224 Yeah, I think we're living in interesting times where it's 1609 01:19:44,232 --> 01:19:49,162 hard to predict what things will be like in 10 years time. 1610 01:19:49,170 --> 01:19:50,732 It's hard to. Yes. 1611 01:19:50,740 --> 01:19:53,952 Is it going to really be different or is it just going to 1612 01:19:53,952 --> 01:19:55,895 stay the same? I have a feeling it will 1613 01:19:55,903 --> 01:19:58,934 genuinely be different. I, I just feel like I I use 1614 01:19:58,942 --> 01:20:01,850 these tools enough and I'm sure you're the same, that I think 1615 01:20:01,858 --> 01:20:04,067 it's more. It's more than hype and anybody 1616 01:20:04,067 --> 01:20:06,790 just says that. I think they need to use these 1617 01:20:06,798 --> 01:20:08,610 tools themselves to see the potential. 1618 01:20:08,618 --> 01:20:11,412 I mean, one of the key things, so we work with customers like 1619 01:20:11,420 --> 01:20:13,612 we really encourage all the executives CEO down, right. 1620 01:20:13,620 --> 01:20:19,580 I mean with a lot of like CTOs, CIOs, but also CEO right, very 1621 01:20:19,580 --> 01:20:22,934 important that they lead by example like embrace these 1622 01:20:22,942 --> 01:20:24,914 tools, right? If you want your word, if you 1623 01:20:24,922 --> 01:20:27,100 believe in this, right. But yeah, like Silicon Valley 1624 01:20:27,108 --> 01:20:29,600 hype cycle is at its all time. That's right. 1625 01:20:29,608 --> 01:20:33,330 Like it is just crazy times. But you know, I I've seen these 1626 01:20:33,330 --> 01:20:35,355 waves before. I you've seen them as well, 1627 01:20:35,363 --> 01:20:38,015 right? Like I do think things do get 1628 01:20:38,023 --> 01:20:42,732 over hyped, but you know, people said cloud was over hyped for a 1629 01:20:42,732 --> 01:20:45,600 long time and actually they were totally wrong. 1630 01:20:45,608 --> 01:20:49,186 It was way under hyped, right? Like go look at a Gartner report 1631 01:20:49,186 --> 01:20:51,490 from like 2010 about like cloud computing, right? 1632 01:20:51,498 --> 01:20:53,690 They probably didn't even have a quadrant, right? 1633 01:20:53,698 --> 01:20:58,116 Like it's just like, but I think my intuition, like I think as 1634 01:20:58,124 --> 01:21:00,800 you said earlier, and it is hard to forecast these things, right? 1635 01:21:00,800 --> 01:21:01,760 Especially like 10 years. 1636 01:21:01,760 --> 01:21:04,360 That was a long time. But from first principles, if 1637 01:21:04,360 --> 01:21:08,440 you could pass the Turing test, that changes a lot of things 1638 01:21:08,440 --> 01:21:10,440 because the way you sort of interact and that so that 1639 01:21:10,440 --> 01:21:15,640 interacting with a human or an AI is similar. 1640 01:21:16,120 --> 01:21:18,040 You can say 1's better than the other, whatever. 1641 01:21:18,080 --> 01:21:22,280 But I think that changes, yeah, how entire organization is 1642 01:21:22,280 --> 01:21:23,840 built, right? Because like, like you said, you 1643 01:21:23,840 --> 01:21:27,840 can manage a bunch of agents instead of people is also maybe 1644 01:21:27,840 --> 01:21:30,400 managing agents is easier than managing people, right? 1645 01:21:30,400 --> 01:21:35,480 Like, you know, like, but it's a, it's a new paradigm. 1646 01:21:35,480 --> 01:21:37,320 And then it's like, of course things are going to be 1647 01:21:37,320 --> 01:21:39,280 overhyped. So if you take AI surrogates as 1648 01:21:39,280 --> 01:21:44,560 an example, you see somebody saying, Hey, you know, you could 1649 01:21:44,560 --> 01:21:46,840 do something 10,000 times faster. 1650 01:21:46,880 --> 01:21:48,680 You know, there's a big. Asterisk there that's sort of 1651 01:21:48,680 --> 01:21:51,520 hidden about like well, but you'd be able to write training 1652 01:21:51,520 --> 01:21:55,844 data and like he gets a 99.9% accuracy, yes, but like even 1653 01:21:55,844 --> 01:21:58,520 more training data and then you know, how does this work 1654 01:21:58,520 --> 01:22:00,120 your organization? Well, then you need this and 1655 01:22:00,120 --> 01:22:07,040 that and so but that's what I call example, like sort of over 1656 01:22:07,040 --> 01:22:09,000 hyping. But then if you look at like the 1657 01:22:09,000 --> 01:22:11,240 real implications, right? Like now all of a sudden you 1658 01:22:11,280 --> 01:22:14,160 take a bunch of tasks that engineers were going to do. 1659 01:22:15,600 --> 01:22:19,280 And if you compress this, like, as we were speaking earlier, 1660 01:22:19,280 --> 01:22:22,840 something from three days into like an hour or less or like a 1661 01:22:22,840 --> 01:22:26,520 second, right? It it's a, it's a sort of like 1662 01:22:26,880 --> 01:22:29,740 there's a lot of implications to how the entire sort of 1663 01:22:29,740 --> 01:22:32,320 ecosystem changes. I don't think you can challenge 1664 01:22:32,320 --> 01:22:33,840 that, right? Like, so you can, you can sort 1665 01:22:33,840 --> 01:22:36,334 of take this one data point and say, like, well, simulations 1666 01:22:36,334 --> 01:22:38,918 are not actually 1000 times faster today versus like 3 1667 01:22:38,918 --> 01:22:41,960 years ago, right? But, you know, on the one hand, 1668 01:22:41,960 --> 01:22:44,960 I don't think you can blame the marketers because like they're 1669 01:22:44,960 --> 01:22:46,720 just doing their job trying to grab your attention. 1670 01:22:46,760 --> 01:22:47,880 It's hard to get somebody's attention. 1671 01:22:48,280 --> 01:22:52,680 Yeah, the like, can something actually be 1000 times faster, 1672 01:22:52,720 --> 01:22:53,950 like given the right conditions? 1673 01:22:53,958 --> 01:22:55,562 Yes, right. And then and then it's like, 1674 01:22:55,570 --> 01:22:58,016 hey, if you implement this in the right way, I think the 1675 01:22:58,024 --> 01:23:00,348 benefits are much more interesting than 1000 times 1676 01:23:00,356 --> 01:23:03,242 faster simulations. It's like now you have designers 1677 01:23:03,242 --> 01:23:06,147 who can get real time physics responses. 1678 01:23:06,155 --> 01:23:08,802 Yeah, right. They're pretty accurate. 1679 01:23:08,810 --> 01:23:13,622 I think you can take this concept of like at a very high 1680 01:23:13,630 --> 01:23:16,324 level with AI. One sort of framework that I 1681 01:23:16,324 --> 01:23:17,960 think is important for organizations to think about is 1682 01:23:17,960 --> 01:23:20,360 like these engineering organizations are highly, highly 1683 01:23:20,360 --> 01:23:22,360 complex. And one of the big challenges is 1684 01:23:22,360 --> 01:23:25,640 the people. So at Boeing, right, you go in 1685 01:23:25,640 --> 01:23:27,480 for like a detailed design review and they're like, oh, we 1686 01:23:27,480 --> 01:23:32,640 got to call up like this world famous expert who's like 78 1687 01:23:32,640 --> 01:23:35,680 years old, right? And they sort of come in and 1688 01:23:35,680 --> 01:23:38,080 they like pontificate and give you advice on like whether this 1689 01:23:38,080 --> 01:23:39,960 is going to work or not. But the intelligence of the 1690 01:23:39,960 --> 01:23:42,560 organization gets sort of lost as the people leave the 1691 01:23:42,560 --> 01:23:44,600 organization. And so like, the problem is if 1692 01:23:44,600 --> 01:23:48,560 you like lose the best people for say, building an airplane, 1693 01:23:49,240 --> 01:23:51,880 if that knowledge was actually just like in their heads and 1694 01:23:51,880 --> 01:23:54,040 like you can kind of see their work, but like you don't really 1695 01:23:54,040 --> 01:23:56,200 kind of understand exactly how they did that. 1696 01:23:56,200 --> 01:23:59,040 And then maybe there was like a aerospace industry goes through 1697 01:23:59,040 --> 01:24:00,720 many cycles. So there's like a time there's 1698 01:24:00,720 --> 01:24:02,080 like 10 years didn't hire anybody. 1699 01:24:02,640 --> 01:24:04,720 So there was no like apprenticeship training, right? 1700 01:24:04,720 --> 01:24:07,908 And so you, you've lost a lot of the ability of the sort of IP 1701 01:24:07,908 --> 01:24:10,811 of like how to actually build great airplanes can be solved 1702 01:24:10,811 --> 01:24:12,956 with AI. AI can kind of like sort of 1703 01:24:12,964 --> 01:24:15,832 again, aggregate a lot of this detailed stuff, synthesize it. 1704 01:24:15,840 --> 01:24:19,057 It's not always going to be perfect or always going to be 1705 01:24:19,065 --> 01:24:21,842 right, but it can do a much better job in a much shorter 1706 01:24:21,850 --> 01:24:24,096 amount of time than any person can really do right. 1707 01:24:24,104 --> 01:24:28,121 And so if you think of AI as ability to do things like that, 1708 01:24:28,121 --> 01:24:31,279 it's pretty incredible. And then like the, if you just 1709 01:24:31,287 --> 01:24:34,022 project forward to even like forget about 10 years, like 6 1710 01:24:34,030 --> 01:24:38,617 months or like 12 months, right? You know, I will say, like when 1711 01:24:38,617 --> 01:24:40,707 I saw ChatGPT the first time, right? 1712 01:24:40,715 --> 01:24:43,492 I'm like, like, interesting toy, right? 1713 01:24:43,500 --> 01:24:47,088 But I didn't actually think at that time, the first time I used 1714 01:24:47,088 --> 01:24:48,920 it so I could just get a change. 1715 01:24:48,920 --> 01:24:51,520 I think it was cool that I could give certain answers, but I 1716 01:24:51,680 --> 01:24:53,600 didn't think it was going to change my like day-to-day 1717 01:24:53,600 --> 01:24:56,760 workflow. And like today, I don't think an 1718 01:24:56,760 --> 01:25:00,200 hour goes by, but I'm not like prompting AI models and and 1719 01:25:00,200 --> 01:25:02,364 running stuff in the background and all kinds of stuff going 1720 01:25:02,364 --> 01:25:05,000 on, right? And so like, that is a new way 1721 01:25:05,000 --> 01:25:06,800 of working. And if you talk to new founders 1722 01:25:06,800 --> 01:25:10,328 that are building new companies, right, that's how they're 1723 01:25:10,328 --> 01:25:14,720 doing it. And they're just like I was, 1724 01:25:16,000 --> 01:25:18,160 there's like the the Claude Code founder, right? 1725 01:25:18,840 --> 01:25:23,304 It's kind of showcasing how he's like does software 1726 01:25:23,304 --> 01:25:25,820 development. He's got like all these 1727 01:25:25,828 --> 01:25:27,650 different agents and all this stuff, right? 1728 01:25:27,658 --> 01:25:30,857 And you know, this Claude Code thing was just a hack day 1729 01:25:30,865 --> 01:25:33,696 project for him like a year ago. And now it's like at I think 400 1730 01:25:33,696 --> 01:25:36,594 million run rate, right? So like this is no joke, right? 1731 01:25:36,594 --> 01:25:38,400 And that's, I think that same 1732 01:25:38,400 --> 01:25:40,820 sort of shift is going to happen in for engineers and 1733 01:25:40,820 --> 01:25:42,680 scientists, right? Like the exact timing of these 1734 01:25:42,680 --> 01:25:46,640 things is always really hard to predict, but same thing with 1735 01:25:46,640 --> 01:25:50,680 like cloud HPC, right? Like I think the the it's easy 1736 01:25:50,680 --> 01:25:53,600 to be right about sort of the secular trends, if you will, at 1737 01:25:53,600 --> 01:25:55,560 least from my perspective, right, where it's like you're 1738 01:25:55,560 --> 01:25:57,640 going to get more processor fragmentation, like Moore's law 1739 01:25:57,640 --> 01:25:59,880 is going to kind of slow down and like got to go to 1740 01:25:59,880 --> 01:26:01,480 specialized processors and all this stuff, right? 1741 01:26:01,480 --> 01:26:03,200 If you want to just get more advanced computing, just all 1742 01:26:03,200 --> 01:26:04,080 these things are going to happen. 1743 01:26:04,200 --> 01:26:07,080 It's very hard to say like at what exact point in time it's 1744 01:26:07,080 --> 01:26:10,680 good this like big shift or whatever and sort of the Overton 1745 01:26:10,680 --> 01:26:12,920 window of what's acceptable to people will shift, right. 1746 01:26:14,360 --> 01:26:16,240 But you could be right about the trends, right? 1747 01:26:16,280 --> 01:26:18,292 And so same thing with AI, I think you can kind of see 1748 01:26:18,292 --> 01:26:21,880 where this is going. And I think then the timing is 1749 01:26:21,880 --> 01:26:24,200 really hard. I would say all the AI 1750 01:26:24,200 --> 01:26:27,280 predictions I would have made in the last 12 months would have 1751 01:26:27,280 --> 01:26:29,880 been like, if they were about something that happened in the 1752 01:26:29,880 --> 01:26:32,520 last 12 months, they would have, I would have predicted them 1753 01:26:32,520 --> 01:26:34,894 over longer time horizons than they actually happened over, 1754 01:26:34,894 --> 01:26:36,570 right? Like, and so if everything is 1755 01:26:36,578 --> 01:26:39,080 just happening much faster, I almost don't even trust my own 1756 01:26:39,088 --> 01:26:40,480 intuition and forecasting too much, right? 1757 01:26:40,488 --> 01:26:43,362 And so like what? Well, what can you do to prepare 1758 01:26:43,362 --> 01:26:47,570 for that is like, I think you got to really lean in to like 1759 01:26:47,578 --> 01:26:49,983 that future. And so like, even if right now, 1760 01:26:49,983 --> 01:26:54,414 like an example is like you can do CAD and CAE modeling through 1761 01:26:54,414 --> 01:26:57,455 natural language prompting through MCP, right? 1762 01:26:57,463 --> 01:27:00,384 So like this concept of model context protocol, think of it 1763 01:27:00,392 --> 01:27:03,625 like an API, right? It's sort of like connect like a 1764 01:27:03,625 --> 01:27:06,664 piece of software with like your, your favorite AI tooling 1765 01:27:06,672 --> 01:27:09,045 and it can interact. So we have a Rescale MCP. 1766 01:27:09,053 --> 01:27:11,381 You can interact with it. You can spin up jobs and do all 1767 01:27:11,381 --> 01:27:13,440 the all the things you'd want to do in the user interface, But 1768 01:27:13,440 --> 01:27:14,785 you can now do it through an LLM. 1769 01:27:14,793 --> 01:27:17,822 But then you can connect to LLM to like a CAD tool, right. 1770 01:27:17,830 --> 01:27:20,660 And you say, hey, like, you know, change the angle of this 1771 01:27:20,668 --> 01:27:22,545 like windshield. You're the author of the 1772 01:27:22,545 --> 01:27:24,780 DrivAerML data set. Like create all these models. 1773 01:27:24,788 --> 01:27:26,105 Like you probably did it manually. 1774 01:27:26,113 --> 01:27:31,769 And so like now you just yeah, so now you just say, hey, you 1775 01:27:31,777 --> 01:27:35,100 know, like building this design space here's like the 1776 01:27:35,108 --> 01:27:37,865 parameters, right, to generate all the CAD. 1777 01:27:37,873 --> 01:27:41,072 It can do that today. Now your CAD designer is going 1778 01:27:41,080 --> 01:27:44,245 to say, Oh yeah, it can do that. But look, look at the 1779 01:27:44,253 --> 01:27:46,390 discontinuity over there. Like because of this thing and 1780 01:27:46,398 --> 01:27:48,700 like whatever. Like it's not good enough yet to 1781 01:27:48,700 --> 01:27:51,384 like 100% replace it. But the whole point is like not 1782 01:27:51,384 --> 01:27:54,238 to 100% replace. The point is now you can 1783 01:27:54,238 --> 01:27:55,895 generate CAD like nobody's business, right? 1784 01:27:55,903 --> 01:27:59,916 Like it's as easy as a prompt that is just going to get way 1785 01:27:59,924 --> 01:28:01,499 better, right? Like, so if, if we're talking 1786 01:28:01,507 --> 01:28:04,744 about like, is this going to be good enough to do the like 1787 01:28:04,752 --> 01:28:07,040 aerospace level carbon fiber layup design work? 1788 01:28:07,048 --> 01:28:11,256 Absolutely at some point, right? Is that in like 1 month or is 1789 01:28:11,264 --> 01:28:13,522 that in 12 months? Or is that in two years? 1790 01:28:13,530 --> 01:28:15,800 I don't know, but it's getting really, really good, right? 1791 01:28:15,808 --> 01:28:18,790 And so those curves are really fast and the adoption curves of 1792 01:28:18,790 --> 01:28:20,920 the of the technology are usually pretty slow in 1793 01:28:20,928 --> 01:28:22,730 enterprise, right? So like just because it's 1794 01:28:22,738 --> 01:28:27,000 possible doesn't mean people adopt it, But you know, these 1795 01:28:27,008 --> 01:28:29,454 industry pressures are real, right? 1796 01:28:29,462 --> 01:28:34,299 Like I, I do think like if you can just kind of do innovation a 1797 01:28:34,299 --> 01:28:38,588 lot faster in automotive, in aerospace, in like semiconductor 1798 01:28:38,588 --> 01:28:43,000 and life sciences, they're all very big R&D spenders, right. 1799 01:28:43,008 --> 01:28:46,115 So if you just get a lot more leverage out of that investment, 1800 01:28:46,115 --> 01:28:49,434 you know, that's a, that's a huge leverage for society, I 1801 01:28:49,442 --> 01:28:52,758 think. So maybe a final question, more 1802 01:28:52,758 --> 01:28:54,950 of a forward-looking question or advice question. 1803 01:28:54,958 --> 01:28:59,642 Given all that we've said, if you are a young founder, OK, 1804 01:28:59,650 --> 01:29:04,592 you're coming out of the first job or a PhD and you, you know, 1805 01:29:04,592 --> 01:29:07,860 you think, what problem can I try and tackle without giving 1806 01:29:07,868 --> 01:29:11,512 away anything that you you're working on so you can't get too 1807 01:29:11,512 --> 01:29:14,704 good around. So I guess what would be like 1808 01:29:14,712 --> 01:29:18,696 the biggest unsolved challenge that you think the next startup 1809 01:29:18,696 --> 01:29:21,454 should try to tackle? Yeah, there's a lot of big 1810 01:29:21,462 --> 01:29:24,610 challenges out there. I think this is going to sound a 1811 01:29:24,610 --> 01:29:26,565 little bit self-serving, but if you're amazing, you can 1812 01:29:26,565 --> 01:29:29,308 always come work at Rescale. My e-mail is joris@rescale.com. 1813 01:29:29,308 --> 01:29:32,080 But I think this look, all the 1814 01:29:32,080 --> 01:29:34,000 things we've discussed, so they're all happening right now, 1815 01:29:34,120 --> 01:29:35,680 right? And so like I say, so where's 1816 01:29:35,680 --> 01:29:37,360 like the sort of next challenge lie? 1817 01:29:38,320 --> 01:29:41,440 I do think it's in this kind of like very buzzwordy word like 1818 01:29:41,440 --> 01:29:43,920 digital twin. But like there is this, you 1819 01:29:43,920 --> 01:29:46,320 know, this concept of digital twin, right, which is the the 1820 01:29:46,320 --> 01:29:49,160 equivalent of like what's happening in the real world in a 1821 01:29:49,160 --> 01:29:53,240 digital form. That concept is super important. 1822 01:29:54,760 --> 01:29:59,120 And this sort of sim-to-real gap, like real being reality, 1823 01:29:59,160 --> 01:30:00,240 right? And the sim being the kind of 1824 01:30:00,240 --> 01:30:03,540 the simulation of that reality, the more you can close that 1825 01:30:03,540 --> 01:30:07,600 gap, the more powerful products you can build, right? 1826 01:30:07,600 --> 01:30:11,720 So like a good example I think is I don't know if you've been 1827 01:30:11,720 --> 01:30:13,160 to San Francisco taking like a Waymo. 1828 01:30:15,000 --> 01:30:16,520 Oh yeah, that freaks me out. Yeah, yeah. 1829 01:30:16,960 --> 01:30:20,400 Yeah, Yeah, I, I think Waymo is like the coolest thing now. 1830 01:30:20,400 --> 01:30:24,160 Why is Waymo so impressive? I remember there were self 1831 01:30:24,160 --> 01:30:29,480 driving vehicles in San Francisco about roughly 10 years 1832 01:30:29,480 --> 01:30:31,440 before. Like Waymo kind of really went 1833 01:30:31,440 --> 01:30:34,880 live for like the sort of private data slash real people, 1834 01:30:34,880 --> 01:30:40,480 right, consumers that last mile of like whether it's regulatory 1835 01:30:40,480 --> 01:30:42,880 or like getting the software to the right level, etcetera. 1836 01:30:44,400 --> 01:30:46,480 Took a long time, the way longer than I expected. 1837 01:30:46,520 --> 01:30:48,680 At that time I would have said like, oh, in a year, if they're 1838 01:30:48,680 --> 01:30:50,463 testing it right now, like in a year, this is going to be 1839 01:30:50,463 --> 01:30:51,840 ready, right? And you sort of know all the 1840 01:30:51,840 --> 01:30:53,720 technology already works. Like you already know, self 1841 01:30:53,720 --> 01:30:55,640 driving kind of works. There's obviously a lot of 1842 01:30:55,640 --> 01:30:58,560 safety and things like that. But that I think that's a great 1843 01:30:58,560 --> 01:31:03,600 example of like Waymo is able to simulate, right, what's 1844 01:31:03,600 --> 01:31:05,840 happening in the real world very effectively. 1845 01:31:06,240 --> 01:31:10,640 Like locally on the edge uses enormous amount of AI, right? 1846 01:31:10,760 --> 01:31:13,080 Enormous amount of training data, right? 1847 01:31:13,320 --> 01:31:17,200 And it, it sort of solves this transportation problem in a way 1848 01:31:17,200 --> 01:31:25,080 that's like awesome, right? Like it's, you know, I and I 1849 01:31:25,080 --> 01:31:28,637 think that's a great example of like what the future looks 1850 01:31:28,637 --> 01:31:30,010 like. Then then you have to ask 1851 01:31:30,018 --> 01:31:32,042 yourself like, OK, well, what are the next set of problems? 1852 01:31:32,050 --> 01:31:34,305 Well, robots is an obvious one, right? 1853 01:31:34,313 --> 01:31:37,010 So like if you can solve this sort of simulation problem for 1854 01:31:37,018 --> 01:31:39,640 robots, I don't know if you've watched these robots like 1855 01:31:39,648 --> 01:31:42,808 folding laundry, but it's like. I kind of wait. 1856 01:31:42,816 --> 01:31:45,782 You know, it's it's yeah, but it's it's not pretty when you 1857 01:31:45,790 --> 01:31:48,344 watch them walk or fold laundry or some of these robot Olympics 1858 01:31:48,344 --> 01:31:51,195 that are going on. We got a ways to go that said, 1859 01:31:51,203 --> 01:31:52,600 like, we all understand physics, right? 1860 01:31:52,608 --> 01:31:54,090 So like, this is actually a solved problem. 1861 01:31:54,098 --> 01:31:57,330 Like we we know how robots should operate in a sort of real 1862 01:31:57,330 --> 01:31:58,750 world, right? And I think NVIDIA's done 1863 01:31:58,750 --> 01:32:01,510 amazing work in, like, sort of, developing the software 1864 01:32:01,510 --> 01:32:03,880 and SDKs for like the virtual world for this, right? 1865 01:32:03,880 --> 01:32:05,880 And so you can sort of train these models in a virtual way, 1866 01:32:06,000 --> 01:32:08,087 but ultimately you want to kind of get that sim-to-real, right? 1867 01:32:08,087 --> 01:32:08,920 Like somebody's going to build 1868 01:32:08,920 --> 01:32:10,440 all these robots. These robots are going to go do 1869 01:32:10,440 --> 01:32:11,720 all the things you don't want to do at home. 1870 01:32:12,280 --> 01:32:14,920 And that's solving those types of problems. 1871 01:32:14,960 --> 01:32:17,720 If you're like sort of have a simulation background, right, is 1872 01:32:18,640 --> 01:32:20,400 I think that's the next frontier, right? 1873 01:32:20,400 --> 01:32:23,240 And you already have this example of Waymo, right? 1874 01:32:23,600 --> 01:32:28,290 And to me, like, I think if you really think closing the 1875 01:32:28,290 --> 01:32:31,000 sim-to- real gap is possible for anything you're passionate 1876 01:32:31,000 --> 01:32:34,360 about, you can do for airplanes, you can do for robots, you can 1877 01:32:34,360 --> 01:32:36,440 do for Earth, right? Like there's this Earth model 1878 01:32:36,680 --> 01:32:38,160 again. NVIDIA has done an amazing job 1879 01:32:38,400 --> 01:32:41,720 developing this weather simulation, you know, 1880 01:32:41,920 --> 01:32:47,000 horrendously difficult problem, but worth solving, right? 1881 01:32:47,000 --> 01:32:50,640 Like if you solve weather prediction, you know, a little 1882 01:32:50,640 --> 01:32:54,680 bit better or even like say 10X better saves a lot of lives, 1883 01:32:54,680 --> 01:32:55,760 right? It makes it makes a big 1884 01:32:55,760 --> 01:32:58,640 difference, right? And yeah, I think those are 1885 01:32:59,040 --> 01:33:02,120 meaningful missions. You know, talk a little bit 1886 01:33:02,120 --> 01:33:04,640 about starting a company. You join a company, Do you, do 1887 01:33:04,640 --> 01:33:06,880 you stay in academia? It's I think you have to really 1888 01:33:06,880 --> 01:33:09,880 think for yourself what what are you motivated by, right? 1889 01:33:10,040 --> 01:33:11,680 Like, like, what do you really want to do, right? 1890 01:33:13,040 --> 01:33:16,800 I personally would encourage people to generally like veer 1891 01:33:16,800 --> 01:33:20,000 towards where they can make the biggest impact and going to 1892 01:33:20,000 --> 01:33:23,680 learn the most right. And so can be a start up, can be 1893 01:33:23,680 --> 01:33:27,360 a big company, could be many different sort of platforms can 1894 01:33:27,360 --> 01:33:32,000 be in university. But I I do think what's what we 1895 01:33:32,000 --> 01:33:35,600 need more of is like the people to really have the courage to 1896 01:33:35,600 --> 01:33:38,600 kind of, I think, solve the problems that are really on the 1897 01:33:38,600 --> 01:33:40,092 edge, but nobody solved before. 1898 01:33:40,100 --> 01:33:43,640 The problem with academia in my view is like I spent a lot of 1899 01:33:43,648 --> 01:33:46,505 time in school, right? A lot too many degrees. 1900 01:33:46,513 --> 01:33:50,880 And it's sort of like, I think the challenge with school is 1901 01:33:50,888 --> 01:33:54,560 that you start competing on a dimension that's that's a little 1902 01:33:54,560 --> 01:33:59,054 bit removed from like the practical reality and you start 1903 01:33:59,054 --> 01:34:00,951 competing. I'm like, you know, can we solve 1904 01:34:00,951 --> 01:34:02,512 this equation more efficiently, whether or not 1905 01:34:02,512 --> 01:34:05,180 Boeing like uses this and adopts this or anybody else 1906 01:34:05,180 --> 01:34:08,480 does, as long as this other really smart person thinks it's 1907 01:34:08,480 --> 01:34:10,040 smart, right? And I get a lot of references, 1908 01:34:10,040 --> 01:34:11,720 like I win the game, so to speak, right? 1909 01:34:11,800 --> 01:34:14,040 But, you know, like it's, I think it's like, I think it's 1910 01:34:14,320 --> 01:34:18,360 Kissinger quote, which is like the battles in academia are so 1911 01:34:18,360 --> 01:34:20,080 fierce because the stakes are so small. 1912 01:34:20,080 --> 01:34:24,400 So, you know, I I think that's a yeah, it's, it's, it's 1913 01:34:24,440 --> 01:34:26,840 unfortunate because you see a lot of really great talent, 1914 01:34:27,040 --> 01:34:28,800 right? And, you know, I was part of 1915 01:34:28,800 --> 01:34:31,360 this as well myself, right? You know, Peter Thiel has a good 1916 01:34:31,360 --> 01:34:34,200 view on this as well. I think where it's like it's 1917 01:34:34,200 --> 01:34:37,680 really elite students who are super smart and they keep 1918 01:34:37,680 --> 01:34:39,080 climbing this sort of academic ladder. 1919 01:34:39,080 --> 01:34:40,000 Why? Because these are like the 1920 01:34:40,000 --> 01:34:42,120 badges you want to collect. I went to Stanford, I went to 1921 01:34:42,120 --> 01:34:44,000 Harvard. I get to this and that right 1922 01:34:44,280 --> 01:34:46,360 have all the sort of certifications, but that's sort 1923 01:34:46,360 --> 01:34:49,080 of like the status game and like signaling to other people, 1924 01:34:49,120 --> 01:34:49,960 right? I think. 1925 01:34:49,960 --> 01:34:53,920 And the problem is, you know, he, what he talks about is that 1926 01:34:53,920 --> 01:34:57,640 like the sort of the in that process, the sort of dreams get 1927 01:34:57,640 --> 01:35:01,560 stomped out of you, right? And that's his view. 1928 01:35:01,840 --> 01:35:03,880 I think that's, that's, that's pretty correct. 1929 01:35:04,280 --> 01:35:06,160 I also think there's a specialization that happens 1930 01:35:06,160 --> 01:35:08,007 because like, OK, you want to be best in the world at 1931 01:35:08,007 --> 01:35:09,840 something, you know, narrow, narrow, narrow, narrow, right? 1932 01:35:10,040 --> 01:35:12,920 But great innovation actually often happens by, you know, 1933 01:35:12,920 --> 01:35:16,000 taking concepts from 1 field, applying it in another, right? 1934 01:35:16,000 --> 01:35:18,360 Like, like it's actually like the DeepMind folks that sort of 1935 01:35:18,360 --> 01:35:21,040 like MeshGraphNets, etcetera. Like that's why we have AI 1936 01:35:21,040 --> 01:35:24,400 surrogates and they're not all like physics experts, right? 1937 01:35:24,440 --> 01:35:26,840 Like, in fact, you often need somebody from a domain who's 1938 01:35:26,840 --> 01:35:29,080 like naive enough to be like, hey, let's just try this thing. 1939 01:35:30,880 --> 01:35:37,560 And so I do think you want people to follow like a mission 1940 01:35:37,560 --> 01:35:40,680 and sort of a like, like the where they're passionate about, 1941 01:35:40,680 --> 01:35:42,160 right? It's like, if you're your 20s, 1942 01:35:42,160 --> 01:35:44,360 like, like you probably have enough life experience to kind 1943 01:35:44,360 --> 01:35:47,920 of know what, what you like. And there's so much more 1944 01:35:47,920 --> 01:35:50,560 possible than people think they are capable of themselves, 1945 01:35:50,920 --> 01:35:52,360 right? Like I personally would have 1946 01:35:52,360 --> 01:35:54,880 never thought when I was like an engineer, simulation engineer 1947 01:35:54,880 --> 01:35:58,720 working at Boeing doing my little thing, right, that I 1948 01:35:58,720 --> 01:36:02,120 could kind of break out, start your own company, get these like 1949 01:36:02,200 --> 01:36:04,480 like maybe they sit across the table from like some of the 1950 01:36:04,480 --> 01:36:08,760 smartest people in Silicon Valley and together kind of like 1951 01:36:08,760 --> 01:36:10,520 it like build an amazing company, right? 1952 01:36:10,520 --> 01:36:14,360 And I think it takes you of course need the ambition, but 1953 01:36:14,360 --> 01:36:17,757 most of all you need the courage and the the willingness to 1954 01:36:17,757 --> 01:36:21,360 kind of persevere, right? Many people give up, like, I 1955 01:36:21,360 --> 01:36:24,443 don't know if you see this, but like, you know, there's a lot 1956 01:36:24,443 --> 01:36:26,600 of people who apply to jobs at Rescale and I see a lot of 1957 01:36:26,600 --> 01:36:30,120 resumes and like you see a lot of this, like, hey, 1 1/2 years 1958 01:36:30,120 --> 01:36:32,400 here, 1 1/2 years there, etcetera. 1959 01:36:32,400 --> 01:36:33,680 Right? I, I think you got to kind of 1960 01:36:33,680 --> 01:36:35,840 find that right, problem that you're really passionate about 1961 01:36:35,840 --> 01:36:42,760 and then like really pursue that with, with all the energy you 1962 01:36:42,760 --> 01:36:44,560 have. There's a good essay written by 1963 01:36:44,560 --> 01:36:48,280 Paul Graham called, it's called great work or something like 1964 01:36:48,280 --> 01:36:50,160 that. It's like how to do great work. 1965 01:36:50,960 --> 01:36:54,096 It's kind of a long essay, so it might take a while to read, 1966 01:36:54,096 --> 01:36:56,960 but one of the things he talks about is like, you know, you 1967 01:36:56,960 --> 01:37:00,170 want to be on the edges, like sort of the he thinks of like 1968 01:37:00,170 --> 01:37:02,691 knowledge as this sort of like tree and these like fractals 1969 01:37:02,691 --> 01:37:04,085 basically. And you're sort of want to be on 1970 01:37:04,085 --> 01:37:06,016 the edge, right? And then you want to kind of see 1971 01:37:06,016 --> 01:37:08,821 where the gaps are. And often I, I think I'm not 1972 01:37:08,829 --> 01:37:12,272 sure if he talks about an essay, but in my view, it's like if you 1973 01:37:12,272 --> 01:37:15,456 take kind of the lessons from a certain field applying in 1974 01:37:15,464 --> 01:37:17,436 another. That's where I've seen like 1975 01:37:17,444 --> 01:37:20,464 amazing breakthroughs like that Rescale and like also beyond 1976 01:37:20,472 --> 01:37:23,240 like in places like Boeing and other places. 1977 01:37:23,248 --> 01:37:26,020 And that requires like the willingness to kind of learn 1978 01:37:26,028 --> 01:37:28,088 these different things, right? And, and sort of apply that 1979 01:37:28,096 --> 01:37:31,350 curiosity in different ways. So, you know, for somebody who 1980 01:37:31,358 --> 01:37:34,980 is just graduating with their, you know, master's or PhD or 1981 01:37:34,988 --> 01:37:37,530 something like I would look to the in short, I would go 1982 01:37:37,538 --> 01:37:40,164 somewhere where I think you're going to learn the most or make 1983 01:37:40,164 --> 01:37:44,659 the biggest impact. What is hard is that there's a 1984 01:37:44,667 --> 01:37:48,264 lot of like societal things that will pressure you to do other 1985 01:37:48,272 --> 01:37:51,020 things, right, Like maybe go work for the company that has 1986 01:37:51,028 --> 01:37:53,515 the brand that your parents will be proud of you for, right? 1987 01:37:53,523 --> 01:37:54,915 Like it's a natural thing, right? 1988 01:37:54,923 --> 01:37:57,845 Like maybe go, you know, like, for example, if you just start a 1989 01:37:57,845 --> 01:38:00,020 company, it's like, well, like now you're unemployed is a 1990 01:38:00,028 --> 01:38:01,406 different view of the same thing, right? 1991 01:38:01,414 --> 01:38:04,962 Like, so I think that some of these things are hard, but but 1992 01:38:04,970 --> 01:38:08,460 you know, I do think people could take much more risks than 1993 01:38:08,460 --> 01:38:10,143 they usually think they can, right? 1994 01:38:10,151 --> 01:38:14,000 And the challenge is a little bit as if you find that out late 1995 01:38:14,000 --> 01:38:15,420 in life, there are less opportunities. 1996 01:38:15,428 --> 01:38:17,860 Like we, we all have just time, right? 1997 01:38:17,868 --> 01:38:20,264 And, and that's the sort of the, the great equalizer. 1998 01:38:20,272 --> 01:38:22,601 You know, people spend a lot of time doing work. 1999 01:38:22,609 --> 01:38:25,720 I certainly do. I think it's like the that time 2000 01:38:25,720 --> 01:38:28,536 that you spent, it should be something really meaningful to 2001 01:38:28,544 --> 01:38:30,230 you. I can't just be a mean student, 2002 01:38:30,230 --> 01:38:31,742 right? Like I love spending time with 2003 01:38:31,750 --> 01:38:33,962 my kids and family and those things too, right? 2004 01:38:33,970 --> 01:38:36,452 But like, if you're a motivated individual, you're probably 2005 01:38:36,460 --> 01:38:37,970 going to spend a lot of time working, right? 2006 01:38:37,978 --> 01:38:41,600 And so that whatever that work is, it should be, you know, I 2007 01:38:41,608 --> 01:38:44,686 wouldn't compromise that too much for, say, like a better 2008 01:38:44,694 --> 01:38:47,216 salary. You're like a mission you're not 2009 01:38:47,216 --> 01:38:50,775 so excited about, right? A lot of big tech companies that 2010 01:38:50,775 --> 01:38:52,800 have, you know, some missions are certainly more interesting 2011 01:38:52,808 --> 01:38:55,922 than others, right? And I think applying yourself to 2012 01:38:55,922 --> 01:38:59,240 kind of make an impact, you know, move society forward is 2013 01:38:59,248 --> 01:39:02,280 super important. Yeah, wise words and thank you 2014 01:39:02,288 --> 01:39:05,137 so much for this. I love this conversation and I 2015 01:39:05,145 --> 01:39:09,579 think what we need to do is schedule in like a few years 2016 01:39:09,587 --> 01:39:14,240 time and see how close we were, whether we're all using AI 2017 01:39:14,248 --> 01:39:16,048 engineers or whether we were wrong. 2018 01:39:16,056 --> 01:39:18,438 I'm sure something completely different will have come around 2019 01:39:18,438 --> 01:39:21,236 that neither of us predicted, but it'd be a good thing to see. 2020 01:39:21,236 --> 01:39:22,360 So with with that, thank you 2021 01:39:22,360 --> 01:39:24,440 Joris, really, really appreciated your time. 2022 01:39:25,200 --> 01:39:26,840 Thank you, Neil. Thanks for the opportunity. 2023 01:39:26,840 --> 01:39:29,720 And yeah, let's let's put it on the calendar and pass it back 2024 01:39:29,720 --> 01:39:31,360 up. Sounds good. 2025 01:39:31,520 --> 01:39:32,600 All right. Cheers. 2026 01:39:32,680 --> 01:39:34,000 Thank you. See you.