1 00:00:00,240 --> 00:00:02,360 Hi, and welcome to the Neil Ashton Podcast. 2 00:00:03,000 --> 00:00:05,880 In each episode, we explained some of the fascinating ways 3 00:00:05,880 --> 00:00:09,080 that science and engineering are changing the world around us. 4 00:00:09,760 --> 00:00:12,720 We talked to leading engineers from elite level sports like 5 00:00:12,800 --> 00:00:16,840 cycling and Formula One to some of the world's top academics to 6 00:00:16,840 --> 00:00:20,960 understand how fluid dynamics, machine learning, supercomputing 7 00:00:21,360 --> 00:00:22,960 are bringing in a new era of discovery. 8 00:00:23,920 --> 00:00:27,040 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 --> 00:00:41,760 episode. Hi, and welcome back to the Neil 12 00:00:41,760 --> 00:00:44,160 Ashton Podcast. So my guest today is Professor 13 00:00:44,160 --> 00:00:48,520 Nathan Kutz, a true pioneer at the intersection of machine 14 00:00:48,520 --> 00:00:51,160 learning, dynamical systems and fluid mechanics. 15 00:00:52,000 --> 00:00:55,680 He, he's really been one of the household names in this field 16 00:00:55,680 --> 00:01:01,206 and has led to, has been part of many of the key sort of 17 00:01:01,206 --> 00:01:06,247 advances and now actually has moved on to play a key role at 18 00:01:06,247 --> 00:01:09,692 one of the major companies in the space at Autodesk 19 00:01:09,692 --> 00:01:11,600 Research. You know, he, he started his 20 00:01:11,608 --> 00:01:12,640 journey at University of Washington. 21 00:01:12,648 --> 00:01:16,585 So he earned a degree in physics and maths there in 1990 and then 22 00:01:16,585 --> 00:01:18,567 did a PhD at Northwestern University. 23 00:01:18,575 --> 00:01:22,272 He returned back to University of Washington and was the chair 24 00:01:22,272 --> 00:01:25,140 of the applied mathematics department and was also co- 25 00:01:25,148 --> 00:01:27,710 directing the AI Institute in Dynamic Systems. 26 00:01:27,718 --> 00:01:32,620 He's probably best known for his work on the dynamic mode 27 00:01:32,628 --> 00:01:37,255 decomposition and DMD and also was a co-author with at his 28 00:01:37,263 --> 00:01:41,812 time, one of his postdocs, Steven Brunton on the SINDy 29 00:01:41,820 --> 00:01:45,332 sparse identification of non linear dynamics, which really 30 00:01:45,340 --> 00:01:52,207 was part of the breakthrough in the push in the mid-2010s, 31 00:01:52,215 --> 00:01:55,360 2016. Well before machine learning for 32 00:01:55,360 --> 00:01:59,024 for science and engineering was was such a household thing 33 00:01:59,024 --> 00:02:01,400 and then such a you know, a key area. 34 00:02:01,408 --> 00:02:06,290 So that arguably they worked on it whilst it was still not clear 35 00:02:06,290 --> 00:02:08,690 how how important this space would be. 36 00:02:08,698 --> 00:02:11,780 And the work that they did really has proved to be a 37 00:02:11,788 --> 00:02:14,260 landmark paper. They also wrote a textbook 38 00:02:14,268 --> 00:02:16,650 Data-driven Science and Engineering, and that was 39 00:02:16,658 --> 00:02:18,875 between Steve Brunton and and Nathan. 40 00:02:18,883 --> 00:02:23,379 But you know, and of course, many other papers, which we'll 41 00:02:23,387 --> 00:02:27,500 link to in the the show notes that that really has been 42 00:02:27,508 --> 00:02:31,299 influential in this field. One of the the reasons why he's 43 00:02:31,299 --> 00:02:35,148 so interesting is that he's isn't just sort of academic. 44 00:02:35,156 --> 00:02:38,380 He actually spent time sabbatical working alongside 45 00:02:38,388 --> 00:02:41,052 Formula One with the McLaren team. 46 00:02:41,060 --> 00:02:45,740 And we talk in this episode about how that was really 47 00:02:45,748 --> 00:02:50,344 influential for his thinking and trying to tackle the more 48 00:02:50,352 --> 00:02:53,624 industrial problems. And that ultimately led to him 49 00:02:53,632 --> 00:02:57,772 moving over and becoming the director of Physics-Informed AI 50 00:02:57,772 --> 00:03:02,020 at Autodesk Research. And I found this conversation 51 00:03:02,028 --> 00:03:05,203 fascinating because we, we sort of went back and forth a little 52 00:03:05,203 --> 00:03:08,068 bit on some of the history and the differences between reduced 53 00:03:08,068 --> 00:03:11,272 order modelling and, and machine learning and then 54 00:03:11,272 --> 00:03:15,920 pivoted a lot to, you know, this, this, this question of is, 55 00:03:15,920 --> 00:03:19,240 can we just spend billions of dollars on data and, and will it 56 00:03:19,240 --> 00:03:22,347 all be good and we have some foundation model or do we really 57 00:03:22,347 --> 00:03:26,235 need to, to have some physics in there to make this more 58 00:03:26,235 --> 00:03:28,560 affordable and achievable and interpretable? 59 00:03:29,480 --> 00:03:33,280 We we also pivoted and talked quite a lot on the, the agents 60 00:03:33,280 --> 00:03:37,880 front and how we see those as being key to distilling 61 00:03:37,880 --> 00:03:42,320 knowledge from, from what is very expert driven processes. 62 00:03:42,720 --> 00:03:45,240 You know, we were saying that it's not good enough just to 63 00:03:45,240 --> 00:03:48,720 have the access to a, to a highly efficient code to 64 00:03:48,720 --> 00:03:52,800 generate data or a, a great machine learning architecture. 65 00:03:52,800 --> 00:03:57,080 You need the human knowledge, the processes that often these 66 00:03:57,080 --> 00:03:59,200 companies had. And we said that distilling this 67 00:03:59,200 --> 00:04:02,840 into agents for those companies could really help them to move 68 00:04:02,840 --> 00:04:04,920 faster. And and we finished off looking 69 00:04:04,920 --> 00:04:08,680 more broadly at where the world's going with this sort of 70 00:04:08,680 --> 00:04:10,880 technology. One of the stand out things that 71 00:04:10,880 --> 00:04:15,062 he said, and I fully agree with him, is in this age, the need 72 00:04:15,062 --> 00:04:17,920 to think big is more important than ever. 73 00:04:17,920 --> 00:04:21,279 Things are moving so fast that actually having that mindset of 74 00:04:21,279 --> 00:04:24,480 really challenging your own thinking is important. 75 00:04:25,240 --> 00:04:28,800 And then the final topic was really more career advice to to 76 00:04:28,800 --> 00:04:32,160 new people in this space and how they should deal with this 77 00:04:32,160 --> 00:04:35,800 changing world of AI and what they should be studying going 78 00:04:35,840 --> 00:04:38,600 forward. As with all these episodes, you 79 00:04:38,600 --> 00:04:41,640 know, I think we could have carried on for many hours. 80 00:04:41,640 --> 00:04:44,203 And I hope I have an opportunity to do that again with Nathan 81 00:04:44,203 --> 00:04:46,680 and speak back to him in a few years and see if some of the 82 00:04:46,680 --> 00:04:50,040 predictions were were correct. But I certainly learned a lot 83 00:04:50,040 --> 00:04:52,960 from this conversation, and I hope that you do too. 84 00:04:52,960 --> 00:04:55,960 So sit back and listen to this conversation with Professor 85 00:04:55,960 --> 00:04:58,920 Nathan Kutz. Well, thank you, Nathan, for for 86 00:04:58,920 --> 00:05:02,080 coming on this. This has been on my sort of wish 87 00:05:02,080 --> 00:05:06,480 list to, to speak to you every time I talk about machine 88 00:05:06,480 --> 00:05:09,880 learning related to engineering, to CFD, to any of these 89 00:05:09,880 --> 00:05:12,040 problems. You know, your name is very high 90 00:05:12,040 --> 00:05:18,040 up the list, some seminal pieces of work and arguably you, 91 00:05:19,040 --> 00:05:21,440 you've been working on this way before. 92 00:05:21,480 --> 00:05:25,640 This was a sort of buzz word topic, you know, a sort of sexy 93 00:05:25,640 --> 00:05:30,200 area to get into. You know, where, where did you 94 00:05:30,200 --> 00:05:31,649 sort of start to get into this? 95 00:05:31,657 --> 00:05:33,015 What, what was your career trajectory? 96 00:05:33,023 --> 00:05:37,826 And when did you see this as like the missing piece to start 97 00:05:37,826 --> 00:05:41,192 using some of these machine learning or reduced order 98 00:05:41,200 --> 00:05:44,621 techniques? Yeah, so I thanks for having me 99 00:05:44,621 --> 00:05:49,268 first of all, Neil to be here. I'm I'm excited to talk to you 100 00:05:49,276 --> 00:05:51,880 as well. And, and, and I would say that 101 00:05:51,888 --> 00:05:56,240 I, I, I suppose I got into it just out of pure interest in 102 00:05:56,248 --> 00:06:01,221 sort of a very accidental way. I was working when I started as 103 00:06:01,221 --> 00:06:04,665 a faculty late 90s, a while ago now. 104 00:06:04,673 --> 00:06:07,330 And that was the beginning of my career. 105 00:06:07,338 --> 00:06:09,956 I was working a lot in atomic and optical physics. 106 00:06:09,964 --> 00:06:13,269 So I was doing theory for mode-locked lasers, lots of 107 00:06:13,277 --> 00:06:16,080 computation, lots of modeling, integration of those two 108 00:06:16,088 --> 00:06:18,194 together, trying to work with experimentalists. 109 00:06:18,202 --> 00:06:25,964 And I really enjoyed that. But around the mid 2000s, what I 110 00:06:25,964 --> 00:06:31,008 found was I, I couldn't get students interested in this 111 00:06:31,016 --> 00:06:33,361 stuff. Like a lot of the students just 112 00:06:33,361 --> 00:06:37,220 did not want to work on these, let's call them harder physics 113 00:06:37,228 --> 00:06:40,842 problems, right, where you had to know some quantum and E&M. 114 00:06:40,850 --> 00:06:44,860 They were very interested at the time in neuroscience, like that 115 00:06:44,860 --> 00:06:50,424 was the hot applied math fields that everybody wanted to do. 116 00:06:50,432 --> 00:06:55,972 And so around 2007, six and seven, I decided, you know, 117 00:06:55,980 --> 00:06:58,029 look, having a hard time getting students. 118 00:06:58,037 --> 00:07:02,107 I really like the work, but there's so many new interesting 119 00:07:02,107 --> 00:07:05,160 areas. And I also had a little bit of, 120 00:07:05,160 --> 00:07:07,926 I guess I'll just say it's an academic crisis. 121 00:07:07,934 --> 00:07:10,262 I think everybody goes through it a little bit where you kind 122 00:07:10,270 --> 00:07:12,020 of get a little bored with what you're doing. 123 00:07:12,028 --> 00:07:17,080 You see that, yes, I could continue this for a long time in 124 00:07:17,080 --> 00:07:21,077 my career or I take the risk and do something really new. 125 00:07:21,085 --> 00:07:23,950 And of course, this is a little scary when you're more of a 126 00:07:23,958 --> 00:07:26,704 senior person because you know, like you're you're like a 127 00:07:26,712 --> 00:07:30,440 beginner again. I remember going to this 128 00:07:30,448 --> 00:07:33,679 neuroscience workshop where clearly every grad student in 129 00:07:33,687 --> 00:07:35,845 the room knew more than me about neuroscience. 130 00:07:35,853 --> 00:07:39,770 But they thought, this odd duck, who's like, well, he seems to 131 00:07:39,770 --> 00:07:42,592 know something like he's not just an illiterate science 132 00:07:42,592 --> 00:07:45,336 person, but on the other hand, doesn't seem to know much 133 00:07:45,336 --> 00:07:46,960 neuroscience. And then, you know, to put 134 00:07:46,960 --> 00:07:50,440 yourself in that, I would say a very vulnerable position was, 135 00:07:51,200 --> 00:07:54,960 was kind of what I did. And I started actually in about 136 00:07:54,960 --> 00:07:59,720 2006 seven to do neuroscience. And then also at that time 137 00:07:59,720 --> 00:08:02,920 seeing some of the data, I thought, hey, there's these data 138 00:08:02,920 --> 00:08:06,080 analysis methods. What if I started integrating 139 00:08:06,080 --> 00:08:09,160 that into sort of building models? 140 00:08:09,960 --> 00:08:15,080 And so that was really the, the initial part of that. 141 00:08:15,080 --> 00:08:18,960 I would say I got into sort of, let's call it broadly, maybe I 142 00:08:18,960 --> 00:08:21,120 wouldn't, maybe you wouldn't call it that now. 143 00:08:21,120 --> 00:08:23,720 It's actually the, the language around what machine learning is 144 00:08:23,720 --> 00:08:26,920 and deep, you know, it's, it's almost all deep learning now. 145 00:08:27,680 --> 00:08:31,800 Let's call, if you go back 2007 and eight, it wasn't neural 146 00:08:31,800 --> 00:08:34,960 networks yet, but I started using a lot of these 147 00:08:34,960 --> 00:08:39,240 methodologies that came out of the data-driven piece. 148 00:08:39,558 --> 00:08:43,240 And really part of the incentive there was I saw that there was 149 00:08:43,240 --> 00:08:45,240 this great opportunity in neuroscience where we didn't 150 00:08:45,240 --> 00:08:47,600 have first principle models, right? 151 00:08:47,600 --> 00:08:51,520 Where you didn't have F equals MA, you didn't have a Maxwell's 152 00:08:51,520 --> 00:08:53,880 equations. You had data and neurons and 153 00:08:53,880 --> 00:08:57,240 people were making up models. And but you're also looking at 154 00:08:57,520 --> 00:09:00,760 trying to build models that, you know, populate population 155 00:09:00,760 --> 00:09:03,400 levels of neurons. And we just started building 156 00:09:03,400 --> 00:09:08,320 these data-driven models back in that time frame, 2007, 2008. 157 00:09:08,800 --> 00:09:12,000 And I thought it was awesome. I just really enjoyed the new 158 00:09:12,000 --> 00:09:14,080 direction. And then I kind of just pivoted 159 00:09:14,080 --> 00:09:17,840 over to there and started working in data-driven modeling. 160 00:09:19,720 --> 00:09:22,674 I started to teach a class on it at the University of 161 00:09:22,674 --> 00:09:23,948 Washington. It was 2008. 162 00:09:23,956 --> 00:09:29,296 So it was very early on. And mostly I just did it because 163 00:09:29,296 --> 00:09:32,502 I, I liked it. I I didn't, I didn't see the 164 00:09:32,510 --> 00:09:34,800 what was coming in 10 years, right. 165 00:09:34,808 --> 00:09:37,458 I mean, it's, it's funny. You know, you can always brag 166 00:09:37,458 --> 00:09:38,960 and. Oh, I saw it. 167 00:09:38,968 --> 00:09:43,186 But I mean, really I just did it because I was interested and I 168 00:09:43,194 --> 00:09:48,096 thought it was very powerful. And I remember going to some of 169 00:09:48,096 --> 00:09:50,399 these optics conferences because I started doing this and 170 00:09:50,399 --> 00:09:53,024 I'm like, wait a minute, we can actually start doing some of 171 00:09:53,024 --> 00:09:56,928 this to model lasers. And I'd show up these optics 172 00:09:56,936 --> 00:09:59,942 conferences and they're like, what is this stuff you're doing? 173 00:09:59,942 --> 00:10:02,200 Like there was like when you're 174 00:10:02,200 --> 00:10:04,920 talking about machine learning integration into the modeling 175 00:10:04,920 --> 00:10:09,320 pipeline at that point, it was it was so foreign to people. 176 00:10:09,320 --> 00:10:12,120 It was like this, you know, this guy's lost his marbles. 177 00:10:12,120 --> 00:10:17,040 He does really falling off the wagon in terms of he's way out 178 00:10:17,040 --> 00:10:19,280 there. And then, you know, you advance 179 00:10:19,280 --> 00:10:23,000 a decade and then everybody is using it, right. 180 00:10:24,120 --> 00:10:25,680 Yeah. I'd love to say I saw that, but 181 00:10:25,680 --> 00:10:29,360 I, I just did it because I liked it and turned out to be 182 00:10:29,360 --> 00:10:31,440 something that everybody started doing. 183 00:10:31,440 --> 00:10:33,480 So that was actually my journey into it. 184 00:10:34,240 --> 00:10:38,320 And then I built from there. And now I'm just in fear of 185 00:10:38,320 --> 00:10:42,320 getting left behind because like these things, they, what they 186 00:10:42,320 --> 00:10:45,080 could do with, you know, is, is unbelievable. 187 00:10:45,080 --> 00:10:48,120 Like they work so much faster than I ever did in my whole 188 00:10:48,160 --> 00:10:52,680 career. That's funny. 189 00:10:52,720 --> 00:10:58,320 So how did you, I mean, obviously the IT looked like 190 00:10:58,320 --> 00:11:03,560 2016 seventeen was like quite big time. 191 00:11:03,560 --> 00:11:05,640 Obviously with with Steve as well. 192 00:11:05,640 --> 00:11:08,000 Like what? How did, how did the years come 193 00:11:08,000 --> 00:11:11,000 up to that sort of seminal piece of work? 194 00:11:11,400 --> 00:11:18,480 Yeah. So I the academic year starting 195 00:11:18,480 --> 00:11:23,200 2012 I had brought on Steve as my postdoc. 196 00:11:23,560 --> 00:11:27,622 He was at Princeton finishing up and I was very fortunate to 197 00:11:27,622 --> 00:11:32,040 get somebody of that quality. Partly it was there was a 2 body 198 00:11:32,040 --> 00:11:35,200 problem. So I, I jointly hired Bing, his 199 00:11:35,200 --> 00:11:39,160 wife with Tom Daniel over in biology and myself as we are. 200 00:11:39,960 --> 00:11:42,880 I didn't have enough money for two postdocs out of money for 1 201 00:11:42,880 --> 00:11:45,240 1/2. So I put half for Bing and got 202 00:11:45,240 --> 00:11:47,800 Tom to do another half and then hired Steve on. 203 00:11:47,800 --> 00:11:50,840 And at that time I was already doing some of the data-driven 204 00:11:50,840 --> 00:11:53,440 modeling, especially things like dynamic mode decomposition. 205 00:11:53,440 --> 00:11:57,960 I had started working on already in 2012/2013. 206 00:11:58,880 --> 00:12:02,080 And so already was this regression framework towards 207 00:12:02,080 --> 00:12:05,800 thinking about dynamical systems and, and, and then it just 208 00:12:05,800 --> 00:12:10,840 turned out that it was just the perfect timing with Steve and 209 00:12:10,840 --> 00:12:16,120 with an amazing group of graduate students and post docs 210 00:12:16,680 --> 00:12:19,080 and also a little bit more open space. 211 00:12:19,080 --> 00:12:22,200 You know, back then it wasn't sort of nowadays it's almost 212 00:12:22,200 --> 00:12:26,280 like there's just so many people working in this area that it's 213 00:12:26,880 --> 00:12:29,760 hard to gain any room. But back then it was still 214 00:12:29,760 --> 00:12:31,550 pretty early on in the sciences. 215 00:12:31,558 --> 00:12:35,095 You know, we have ImageNet, right 2014 that hadn't quite 216 00:12:35,103 --> 00:12:37,880 made its transition over to the sciences. 217 00:12:37,888 --> 00:12:42,092 I think that would more like 2018, '19, '20 like where they 218 00:12:42,092 --> 00:12:46,390 were really getting into it. So we were early on had some 219 00:12:46,398 --> 00:12:50,576 free space there and also had less rules I guess. 220 00:12:50,584 --> 00:12:54,720 And we just started doing some stuff there. 221 00:12:54,728 --> 00:12:58,949 And I think obviously SINDy came out of that time period and some 222 00:12:58,949 --> 00:13:02,484 PDE-FIND, there's the dynamic mode decomposition work and then 223 00:13:02,484 --> 00:13:06,561 tying all that together with what was happening also in 224 00:13:06,561 --> 00:13:09,550 reduced order modeling. So it just was a really 225 00:13:09,558 --> 00:13:12,136 fantastic time. And then Steve then took a 226 00:13:12,144 --> 00:13:14,760 faculty position there in mechanical engineering and then 227 00:13:14,760 --> 00:13:19,544 we continue to have a really great partnership with our 228 00:13:19,552 --> 00:13:23,275 students and postdocs. And it was incredibly productive 229 00:13:23,275 --> 00:13:26,384 in terms of starting to build, you know, these data-driven 230 00:13:26,392 --> 00:13:30,640 models and trying to bring in machine learning overall into 231 00:13:30,648 --> 00:13:33,275 the standard dynamical systems framework. 232 00:13:33,283 --> 00:13:37,622 So maybe for people who are not so familiar, how would you, how 233 00:13:37,622 --> 00:13:40,736 would you describe SINDy and the sort of POD methods? 234 00:13:40,744 --> 00:13:45,119 You know what, what was the what was the core aim or thing that 235 00:13:45,127 --> 00:13:48,006 you were trying to solve? Yeah. 236 00:13:48,014 --> 00:13:51,872 So I think, I think there's there's two different aspects 237 00:13:51,880 --> 00:13:55,219 here. So, so I always think about 238 00:13:55,227 --> 00:14:00,492 physics as sort of, if you go back to like every engineer, 239 00:14:00,500 --> 00:14:03,840 they got sort of a classic except for computer scientists. 240 00:14:03,840 --> 00:14:05,160 If you want to consider them 241 00:14:05,160 --> 00:14:07,840 engineers, let's they're a little different brand of 242 00:14:07,920 --> 00:14:11,048 engineer, but almost every other engineer had to take statics 243 00:14:11,048 --> 00:14:14,360 and dynamics. And then statics and dynamics. 244 00:14:14,360 --> 00:14:17,160 There was only one thing you really did, which is you drew a 245 00:14:17,160 --> 00:14:21,200 free body diagram. And then once you drew the right 246 00:14:21,200 --> 00:14:25,520 free body diagram, that was really almost every homework set 247 00:14:25,520 --> 00:14:27,880 was just draw the right free body diagram. 248 00:14:27,880 --> 00:14:31,240 Because once you draw the right free body diagram, you could 249 00:14:31,240 --> 00:14:34,760 actually either sum of forces equals 0, that's statics sum of 250 00:14:34,760 --> 00:14:39,400 forces equals not 0. But if you didn't draw the right 251 00:14:39,400 --> 00:14:42,760 free body diagram, it was going to be a really hard solve. 252 00:14:44,640 --> 00:14:48,240 And the way I think about that in context of what we are trying 253 00:14:48,240 --> 00:14:52,160 to do, like even with POD or these, you know, even what we do 254 00:14:52,160 --> 00:14:55,960 now with encoders is can I find the right representation or the 255 00:14:55,960 --> 00:14:59,480 right coordinate system in which to express my dynamics? 256 00:15:01,040 --> 00:15:05,767 And so, so that was already sort of a philosophical idea of 257 00:15:05,767 --> 00:15:09,094 like, hey, we have all this data, you know, it's high 258 00:15:09,094 --> 00:15:11,480 dimensional, it looks crazy. But at the end of the day that 259 00:15:11,480 --> 00:15:15,040 we can find these some kind of embedding where it's actually 260 00:15:15,040 --> 00:15:18,160 looks kind of simple maybe, right. 261 00:15:18,560 --> 00:15:21,480 But then what's happening in that space is now you have to 262 00:15:21,480 --> 00:15:23,840 have some kind of dynamical system relationships. 263 00:15:23,840 --> 00:15:28,760 And so the SINDy part really comes from this idea that if we 264 00:15:28,760 --> 00:15:33,120 look at our physics models, all the physics models that we have 265 00:15:33,120 --> 00:15:38,960 are really essentially expressions of relationships 266 00:15:38,960 --> 00:15:42,400 among derivatives, time and space, right? 267 00:15:42,400 --> 00:15:49,160 And it's in most remarkable compression of knowledge that 268 00:15:49,160 --> 00:15:52,960 we, I think we have ever developed as humans, right? 269 00:15:52,960 --> 00:15:54,720 You know, you write down Maxwell's equations. 270 00:15:54,720 --> 00:15:57,120 You have this, you fit it on the T-shirt, right? 271 00:15:57,120 --> 00:16:00,160 You can order online on Amazon this afternoon, you get it 272 00:16:00,160 --> 00:16:01,760 right. And you have those Maxwell's 273 00:16:01,760 --> 00:16:04,920 equations on your T-shirt. And what's amazing about it is 274 00:16:05,200 --> 00:16:10,240 it models, you know, all of our Wi-Fi signals like coming, you 275 00:16:10,240 --> 00:16:15,560 know, radar, you know, radio waves, the diversity of what 276 00:16:15,560 --> 00:16:20,160 you're modeling in this compact representation, which is, is 277 00:16:20,160 --> 00:16:22,440 astounding. Same thing with how we are able 278 00:16:22,440 --> 00:16:25,040 to get, you know, when think about quantum mechanics, it's 279 00:16:25,040 --> 00:16:26,960 like, you know, 3 terms, right? 280 00:16:26,968 --> 00:16:30,360 You have your dispersion term, your potential and, and your 281 00:16:30,368 --> 00:16:33,678 time derivative. And it's, so I think the 282 00:16:33,686 --> 00:16:39,140 motivation for SINDy was this idea that we've had tremendous 283 00:16:39,148 --> 00:16:45,601 success in history with these compact representations of 284 00:16:45,609 --> 00:16:50,424 relationships among derivatives. And so the SINDy algorithm goes 285 00:16:50,424 --> 00:16:55,230 directly after that says, well, there's lots of derivatives and 286 00:16:55,230 --> 00:16:59,696 derivatives relationships. When you look at this data, how 287 00:16:59,696 --> 00:17:02,970 few of these derivative relationships can you use that 288 00:17:02,978 --> 00:17:07,019 actually fit the data? And that's where this idea of 289 00:17:07,027 --> 00:17:10,487 having a library of potential candidates and, and then 290 00:17:10,494 --> 00:17:15,656 sparsely regressed to just a few terms came out of. 291 00:17:15,664 --> 00:17:19,412 And it's kind of a philosophical thing. 292 00:17:19,420 --> 00:17:21,412 So I think like physicists like this idea. 293 00:17:21,420 --> 00:17:24,020 It's not clear that computer scientists like this idea. 294 00:17:24,028 --> 00:17:27,028 Like computer scientists might, I've heard this directly. 295 00:17:27,036 --> 00:17:30,835 They feel like, well, that's your, that's your 20th century 296 00:17:30,843 --> 00:17:33,160 crutch of physics. Like, that's because that's 297 00:17:33,168 --> 00:17:36,584 what, you know what, let's let the computer figure out. 298 00:17:36,592 --> 00:17:40,682 Right. So it's still a tension point, 299 00:17:40,690 --> 00:17:46,923 but as we talk further about some of the where physics AI is 300 00:17:46,923 --> 00:17:51,700 going, I think one of the things that's really incredibly 301 00:17:51,700 --> 00:17:56,890 valuable to consider is that governing equations so far, I, 302 00:17:56,898 --> 00:18:00,129 I, I think this is true. I mean, it's hard to, you know, 303 00:18:00,129 --> 00:18:02,355 these are statements and of opinions, I guess. 304 00:18:02,363 --> 00:18:06,336 But, but really historically, we can see that they've been the 305 00:18:06,344 --> 00:18:09,068 most incredible engines of extrapolation in predictions 306 00:18:09,076 --> 00:18:13,256 where I've never had data about something that might happen. 307 00:18:13,264 --> 00:18:17,040 You know, if I go over there in parameter space, my model 308 00:18:17,048 --> 00:18:21,150 predicts like this bifurcation, like the change of the system. 309 00:18:21,158 --> 00:18:25,200 And then you can start doing an experiment over there and 310 00:18:25,208 --> 00:18:28,560 validate that, right? Whereas machine learning, it's 311 00:18:28,568 --> 00:18:32,362 very difficult to predict outside of if you don't have 312 00:18:32,370 --> 00:18:35,640 data there, like you're just not going to get it right. 313 00:18:35,648 --> 00:18:39,680 So governing equations still, at least in my heart, hold massive 314 00:18:39,680 --> 00:18:42,520 value because of this extrapolation capability that 315 00:18:42,528 --> 00:18:46,464 we, I think are going to still need in the future. 316 00:18:46,472 --> 00:18:51,240 Yeah, that that's a very good transition point, I guess. 317 00:18:51,248 --> 00:18:54,859 Well, first of all, I'd be interested to how you would 318 00:18:54,867 --> 00:18:59,544 describe before we get on to the maybe today's debate on machine 319 00:18:59,544 --> 00:19:04,600 learning methods and surrogates and, you know, the influence of 320 00:19:04,600 --> 00:19:08,588 physics versus data-driven, some more skeptics, you know, 321 00:19:08,588 --> 00:19:13,396 we'll, we'll sort of say, ah, this whole buzz now about 322 00:19:13,396 --> 00:19:15,240 machine. It's just basically reduced 323 00:19:15,240 --> 00:19:17,160 order modelling. We've been doing it, you know, 324 00:19:17,480 --> 00:19:21,800 for for ages. How would you define the 325 00:19:21,800 --> 00:19:26,440 differences and the similarities between sort of those earlier 326 00:19:26,440 --> 00:19:30,680 reduced order modelling sort of techniques and today with the 327 00:19:30,680 --> 00:19:35,120 more sort of transformer based neural networks, what we would 328 00:19:35,120 --> 00:19:38,680 call sort of machine learning in a very broad term? 329 00:19:39,880 --> 00:19:45,840 Yeah, so early days of reduced order modeling, I think there 330 00:19:45,840 --> 00:19:49,240 was still this idea that these POD modes, you know you do the 331 00:19:49,240 --> 00:19:51,720 SVD, you have these structures that come out. 332 00:19:52,080 --> 00:19:57,840 I think it's still allowed for some flexibility in the idea of 333 00:19:57,840 --> 00:20:02,040 an of interpretation, right. So you you would have things 334 00:20:02,040 --> 00:20:06,160 that were coming out that you would argue made sense, like so 335 00:20:06,160 --> 00:20:09,160 for instance even. POD modes of flow around the 336 00:20:09,160 --> 00:20:13,320 cylinder, you look at it and go like I can explain what that is 337 00:20:13,520 --> 00:20:17,960 and and there was something very comforting about that like to 338 00:20:17,960 --> 00:20:21,120 have sort of in some sense a basis. 339 00:20:21,120 --> 00:20:26,040 And so I think for us, if you look at the 20th century applied 340 00:20:26,040 --> 00:20:30,480 math literature and what the impact was, we did a lot of 341 00:20:30,480 --> 00:20:35,760 expansions and bases or a transform was a workhorse. 342 00:20:35,760 --> 00:20:38,120 We could say look in signal processing, right? 343 00:20:38,120 --> 00:20:42,484 Like just do a Fourier transform process everything in the 344 00:20:42,484 --> 00:20:44,880 signal domain come back. And we had this idea of 345 00:20:44,880 --> 00:20:47,600 interpretability built in because we kind of knew what 346 00:20:47,600 --> 00:20:51,120 cosines and sines were. We followed that up even with 347 00:20:51,120 --> 00:20:54,880 wavelets. So we and even special functions 348 00:20:54,880 --> 00:20:57,160 which were sort of like in mathematical physics, the 349 00:20:57,160 --> 00:21:01,200 foundation of the 1950s and 60s for doing analysis of systems, 350 00:21:02,040 --> 00:21:05,080 it was this idea that there is a basis functions. 351 00:21:05,440 --> 00:21:08,320 In other words, again, I would call this a nice coordinate 352 00:21:08,320 --> 00:21:10,360 system that's sort of interpretable to you. 353 00:21:10,360 --> 00:21:13,080 So you can project everything into this, do your work in 354 00:21:13,080 --> 00:21:17,590 there, you feel comfortable with it because it's you interpret 355 00:21:17,590 --> 00:21:20,760 it and then you can come back and do your and your analysis. 356 00:21:20,760 --> 00:21:22,720 And by the way, we're very successful with this, right? 357 00:21:23,840 --> 00:21:26,760 And largely we constrained ourselves in the Fifties, 60s 358 00:21:26,760 --> 00:21:32,720 and 70s, even up to the 80s to linear problems, because then we 359 00:21:32,720 --> 00:21:36,320 could use superposition, we could have solutions which were 360 00:21:36,320 --> 00:21:38,120 linear combinations of these things. 361 00:21:39,120 --> 00:21:42,280 And then I think at the end of the 80s, of course, in the early 362 00:21:42,280 --> 00:21:45,120 90s, we had the first computational revolution, which 363 00:21:45,120 --> 00:21:49,920 was, oh, I could just throw that PDE on the computer and 364 00:21:49,920 --> 00:21:51,739 discretize and start simulating. 365 00:21:51,747 --> 00:21:57,408 And so, so I feel like when I went to grad school, I was in 366 00:21:57,416 --> 00:22:01,132 grad school in 1990 to 94 and I feel like I was the first 367 00:22:01,140 --> 00:22:03,976 generation where there was an expectation you're going to 368 00:22:03,984 --> 00:22:07,606 simulate PDEs and ODEs in your thesis somewhere. 369 00:22:07,614 --> 00:22:14,140 At that point we had now, you know, desktop computers in the 370 00:22:14,148 --> 00:22:17,080 lab. It wasn't like there's some 371 00:22:17,088 --> 00:22:21,000 central computer, a DEC machine somewhere in a court, you know, 372 00:22:21,000 --> 00:22:23,674 that you run jobs on. This is now like, oh, it's just 373 00:22:23,674 --> 00:22:25,840 sitting here in the office and you're interacting and 374 00:22:25,848 --> 00:22:30,265 programming with it. And so for the first time, we 375 00:22:30,273 --> 00:22:34,120 are actually starting able to solve nonlinear PDEs because 376 00:22:34,128 --> 00:22:37,458 we could just numerically simulate them. 377 00:22:37,466 --> 00:22:40,092 And of course, it's very interesting because I think the 378 00:22:40,092 --> 00:22:42,600 faculty back then were like, you know, these kids, they don't 379 00:22:42,600 --> 00:22:45,560 know what they're doing, understand physics yet they're 380 00:22:45,568 --> 00:22:47,997 doing, you know, they're claiming all this stuff. 381 00:22:48,005 --> 00:22:51,520 And, and of course, it's sort of true, right, because they were 382 00:22:51,528 --> 00:22:54,340 used to solving things, really understanding the problem at 383 00:22:54,348 --> 00:22:58,160 some fundamental level and especially doing lots of 384 00:22:58,168 --> 00:23:01,760 asymptotic reductions in two corners of parameter space where 385 00:23:01,760 --> 00:23:04,520 you could linearize and say quite a bit, right? 386 00:23:04,528 --> 00:23:07,375 That's, that's almost the whole fluid mechanics mantra. 387 00:23:07,383 --> 00:23:12,144 It's like, look over here, we have JFM to tell us what's going 388 00:23:12,144 --> 00:23:17,202 over there on, right? But then we had that simulation 389 00:23:17,202 --> 00:23:21,752 capability and then and then we start saying, OK, but so we 390 00:23:21,760 --> 00:23:24,812 still had some interpretation like I'm expanding in Fourier 391 00:23:24,812 --> 00:23:27,680 basis or I'm doing finite differences, finite elements. 392 00:23:27,688 --> 00:23:30,755 And then, and then from there, from those simulations, then we 393 00:23:30,755 --> 00:23:32,961 started a new basis, which was the POD, right? 394 00:23:32,969 --> 00:23:36,227 Which is like, oh, let's just build basis directly from the 395 00:23:36,235 --> 00:23:38,870 SVD. So we still felt it was very 396 00:23:38,878 --> 00:23:42,480 morally equivalent to what we were kind of doing for. 397 00:23:42,488 --> 00:23:46,405 And then the, the transformation now is sometimes it's just 398 00:23:46,413 --> 00:23:50,087 really hard to get your head around what happened in this 399 00:23:50,095 --> 00:23:52,300 thing. Like I transformer and then I 400 00:23:52,308 --> 00:23:56,551 have an attention layer and I got right, I got, I got this 401 00:23:56,559 --> 00:23:58,500 decoder that's then being modulate. 402 00:23:58,508 --> 00:24:01,245 I mean, right. We, we could see some of the 403 00:24:01,253 --> 00:24:03,860 sophistication and some of the, some of the structures are kind 404 00:24:03,860 --> 00:24:07,528 of simple, but more and more they're becoming fairly complex 405 00:24:07,528 --> 00:24:12,552 and it's really hard for us to decipher what's happening there. 406 00:24:12,552 --> 00:24:14,400 But you can't argue often with 407 00:24:14,840 --> 00:24:17,680 how well it works, right? You get the results and you're 408 00:24:17,680 --> 00:24:19,720 going like, yeah, but I can't beat this. 409 00:24:21,200 --> 00:24:24,280 So I have to, you know, I want to use it. 410 00:24:24,400 --> 00:24:27,960 And by the way, I started to appreciate computer scientists a 411 00:24:27,960 --> 00:24:31,440 little bit, quite a bit. Because at first, you know, I 412 00:24:31,440 --> 00:24:34,280 was like that old grumpy professor, which is like these 413 00:24:34,280 --> 00:24:36,560 computer science kids, they don't know any physics. 414 00:24:36,560 --> 00:24:40,480 They just go do stuff and they get a, they get a score and they 415 00:24:40,480 --> 00:24:43,560 got a cross validation score and they move on with their lives, 416 00:24:43,560 --> 00:24:48,520 right. The flip side now is I have 417 00:24:48,520 --> 00:24:55,720 grown to absolutely love the fearlessness of a, you're a 418 00:24:55,720 --> 00:24:59,480 science grad student. They just like try stuff not 419 00:24:59,480 --> 00:25:03,040 encumbered by all this physics knowledge. 420 00:25:03,040 --> 00:25:07,440 Like, you know, like if I look at this typical physics student 421 00:25:07,440 --> 00:25:10,920 or engineering or applied math, they kind of are, they're 422 00:25:10,920 --> 00:25:13,520 prejudiced by their what they learned. 423 00:25:13,520 --> 00:25:17,230 So they're like, they would try to make smart things 424 00:25:17,230 --> 00:25:19,648 structures, whereas the computer scientists are just like, just 425 00:25:19,648 --> 00:25:21,360 try stuff. What if we did this? 426 00:25:21,368 --> 00:25:25,628 What if we did this? And and that fearlessness has 427 00:25:25,636 --> 00:25:28,490 led them to unbelievable results. 428 00:25:28,498 --> 00:25:32,820 And I think it's going to be left to the applied 429 00:25:32,828 --> 00:25:36,652 mathematicians and engineers and theory people to kind of pick up 430 00:25:36,652 --> 00:25:38,680 the pieces to try to understand. 431 00:25:38,680 --> 00:25:41,320 OK, but what did you what OK in here? 432 00:25:41,320 --> 00:25:44,143 I we can now got to dissect this to see what you actually 433 00:25:44,143 --> 00:25:45,960 learned here, because somehow it is working. 434 00:25:45,960 --> 00:25:47,280 I'm going to give you credit for that. 435 00:25:47,280 --> 00:25:50,560 But I'd really love to understand what this thing's 436 00:25:50,840 --> 00:25:52,880 actually learning about the physics, right? 437 00:25:52,880 --> 00:25:55,160 And that and I think that's kind of what big next steps that's 438 00:25:55,160 --> 00:25:56,760 going to happen in the community. 439 00:25:56,760 --> 00:26:00,080 So you'd say almost the and I've I've heard similar things. 440 00:26:00,080 --> 00:26:06,040 The interpretability is the key bit like how does this work? 441 00:26:06,040 --> 00:26:11,080 Why does it work? What is it that is the knob that 442 00:26:11,080 --> 00:26:13,720 has changed that you know, how do you do? 443 00:26:13,800 --> 00:26:19,240 And I guess like you say, it's all good when it works, but when 444 00:26:19,240 --> 00:26:23,120 it doesn't work, it's like why didn't it work is the question. 445 00:26:23,160 --> 00:26:25,680 Yeah. And at least in our standard 446 00:26:25,680 --> 00:26:28,880 physics modelling, right, we always had some recourse to like 447 00:26:29,320 --> 00:26:32,760 there was just a few parameters we could sort of, we could 448 00:26:32,760 --> 00:26:36,200 actually diagnose why it broke or, or, or went through a 449 00:26:36,200 --> 00:26:38,360 bifurcation or there was a transition. 450 00:26:38,640 --> 00:26:43,700 Like we had a way understanding what would might happen where 451 00:26:43,700 --> 00:26:47,840 in these things. It's a little bit like, I don't 452 00:26:47,840 --> 00:26:50,440 know what kind of car you drive, but like when something breaks 453 00:26:50,440 --> 00:26:55,280 there, you just automatically have to go to a mechanic because 454 00:26:55,280 --> 00:26:58,280 he's caught. The cars are so complex now that 455 00:26:58,280 --> 00:27:01,640 you have to have specialist fix it where, you know, I think, 456 00:27:01,640 --> 00:27:04,832 you know, when I was a kid, I think there's a lot of people 457 00:27:04,832 --> 00:27:06,784 that fix their own cars, right? 458 00:27:06,792 --> 00:27:10,687 It was a little bit more interpretable what was happening 459 00:27:10,687 --> 00:27:14,240 in that engine. But now it's like, no, no, no, 460 00:27:14,248 --> 00:27:16,956 that everything's run off a bunch of electronic components 461 00:27:16,964 --> 00:27:21,360 with all these different pieces. And I think of it like that, 462 00:27:21,368 --> 00:27:23,522 like that. And when they break, you just 463 00:27:23,530 --> 00:27:25,647 have no idea and started fixing it. 464 00:27:25,655 --> 00:27:28,748 Maybe you just train another neural network or the different 465 00:27:28,748 --> 00:27:30,530 structure that doesn't break there. 466 00:27:30,538 --> 00:27:34,672 Yeah, Yeah, Yeah. I mean, how, how do you see then 467 00:27:34,672 --> 00:27:37,462 today because that that just seemed to be the challenge of 468 00:27:37,470 --> 00:27:41,912 our time, doesn't it that you are, I mean, I guess out of 469 00:27:41,920 --> 00:27:44,990 distribution versus in distribution the train on a 470 00:27:44,998 --> 00:27:50,085 bunch of cars. Is it, is it just the case that 471 00:27:50,085 --> 00:27:54,320 that is a essentially not unsolvable, but a problem that 472 00:27:54,328 --> 00:27:58,172 can only really be addressed just by having more data that 473 00:27:58,180 --> 00:28:00,250 makes that out of distribution in distribution. 474 00:28:00,258 --> 00:28:03,552 And so the solution is just keep building bigger and bigger 475 00:28:03,560 --> 00:28:10,274 models with more and more data? Or do you feel that that is 476 00:28:10,282 --> 00:28:13,860 essentially inefficient and and not a scalable method and 477 00:28:13,868 --> 00:28:19,444 therefore you need some ability to, as you say you know, use 478 00:28:19,452 --> 00:28:24,136 some modeling knowledge or physics knowledge to be able to 479 00:28:24,136 --> 00:28:30,323 model past the data? Yeah, I OK, so this question is 480 00:28:30,323 --> 00:28:36,532 hard to answer for, for for one reason, which is not scientific 481 00:28:36,532 --> 00:28:42,915 at all, which is just money. When you look at this product, 482 00:28:42,923 --> 00:28:46,975 you know, Project Prometheus for instance, that Bezos is involved 483 00:28:46,975 --> 00:28:51,180 with and you look at their what 6.2 billion. 484 00:28:51,188 --> 00:28:58,108 Now at some point you ask questions like it's OK, even now 485 00:28:58,108 --> 00:29:00,360 being at Autodesk. So first of all, let's if I, if 486 00:29:00,360 --> 00:29:04,442 I go back to when I was a faculty, the ability to get data 487 00:29:04,442 --> 00:29:08,440 like really say we need to collect this data. 488 00:29:08,448 --> 00:29:11,440 It's like, oh, wait, you're going to have to write some 489 00:29:11,448 --> 00:29:12,749 massive grants. There's probably almost no way 490 00:29:12,757 --> 00:29:15,160 you're going to get the kind of grant levels you need to collect 491 00:29:15,160 --> 00:29:17,920 the data you would need. So you have, you're already 492 00:29:17,928 --> 00:29:21,144 constrained saying I've got to outsmart this and got to figure 493 00:29:21,144 --> 00:29:25,964 out how to do this. Now you move to a company and 494 00:29:25,972 --> 00:29:29,642 now I'm at Autodesk. But even then you're like, yeah, 495 00:29:29,642 --> 00:29:34,420 but you it's, it's not clear. There's a few companies, there's 496 00:29:34,420 --> 00:29:37,235 a there's the trillion dollar companies, right? 497 00:29:37,243 --> 00:29:43,060 That potentially, if they desire to do so, could potentially set 498 00:29:43,060 --> 00:29:47,306 up the architecture to just collect the data, spend the 499 00:29:47,314 --> 00:29:51,704 billions to do things like this and and do that. 500 00:29:51,712 --> 00:29:56,543 Now, that's a brute force approach, but I think the win is 501 00:29:56,543 --> 00:30:00,600 so big that they might just say, yeah, we could spend $10 502 00:30:00,600 --> 00:30:03,110 billion because it's going to make us a trillion. 503 00:30:03,118 --> 00:30:07,456 Like very few places that can do it. 504 00:30:07,464 --> 00:30:10,020 Like, but even Project Prometheus might have the kind 505 00:30:10,028 --> 00:30:15,056 of cash resources because my normal answer would be like, it 506 00:30:15,056 --> 00:30:21,660 is completely, I think, intractable to get the kind of 507 00:30:21,668 --> 00:30:24,980 data we need to do something with the science problems we 508 00:30:24,988 --> 00:30:28,560 need. But when you have 6.2 billion, 509 00:30:28,568 --> 00:30:32,460 right, and maybe more, I don't know, like maybe you go like, 510 00:30:32,468 --> 00:30:36,570 you know, we can do it. Like we're going to spend $3 511 00:30:36,578 --> 00:30:38,774 billion to collect the data, right? 512 00:30:38,782 --> 00:30:41,426 That's higher investment in the in the data. 513 00:30:41,434 --> 00:30:45,060 It isn't is that money is to get the data and then now we can 514 00:30:45,068 --> 00:30:48,139 build whatever the foundation model we need or so. 515 00:30:48,147 --> 00:30:51,414 So that one, I think the jury's still out. 516 00:30:51,422 --> 00:30:54,720 I so, you know, it goes right back to, you know, Richard 517 00:30:54,728 --> 00:30:58,780 Sutton's right, the the bitter lesson issue, which is so 518 00:30:58,788 --> 00:31:02,822 Sutton, I think is right. If you have enough data, right, 519 00:31:02,822 --> 00:31:07,257 you really, if you can collect enough data in an area, it's so 520 00:31:07,257 --> 00:31:11,156 far what we've seen, it's almost these machine learning, 521 00:31:11,156 --> 00:31:14,880 deep learning algorithms and structures are, are basically 522 00:31:14,888 --> 00:31:18,960 unbeatable. But the question is, can you 523 00:31:18,968 --> 00:31:21,558 really do this? I, I love also Max Welling wrote 524 00:31:21,558 --> 00:31:23,468 a response to this, which was well, yeah. 525 00:31:23,476 --> 00:31:27,940 But a lot of times what we're trying to do scientifically is 526 00:31:27,948 --> 00:31:30,903 extrapolate. We're trying to build a new 527 00:31:30,911 --> 00:31:33,853 technology. So even if I collect all this 528 00:31:33,861 --> 00:31:37,500 data, the the goal is to give me a great direction of like 529 00:31:37,508 --> 00:31:39,756 actually, I think the future technologies over there, I'd 530 00:31:39,764 --> 00:31:43,070 have no data there Now. Maybe I could set up a pipeline, 531 00:31:43,070 --> 00:31:46,302 have enough money to to build the scaffolding for the data to 532 00:31:46,302 --> 00:31:50,760 collect on the way out there. But, you know, this is what I 533 00:31:50,768 --> 00:31:54,440 learned in my time at McLaren too, is the answer. 534 00:31:54,448 --> 00:31:59,098 The fast car is an extrapolation that's far away from where 535 00:31:59,106 --> 00:32:01,735 you're starting. Actually, far away is a little 536 00:32:01,743 --> 00:32:04,707 bit of exaggeration. Tenths of second. 537 00:32:04,715 --> 00:32:08,600 Yeah, well, it feels far away when you're trying to like, you 538 00:32:08,600 --> 00:32:10,700 know, the difference between winning the world title. 539 00:32:10,708 --> 00:32:13,516 Like, you know, which McLaren did while I was the time I was 540 00:32:13,524 --> 00:32:16,765 there, they were like had them MCL38/39, which were winners. 541 00:32:16,773 --> 00:32:21,422 You know, you're, you're working at the margins of these updates, 542 00:32:21,422 --> 00:32:24,760 they're trying to get you 2/10 of a second, a 10th of a second, 543 00:32:24,760 --> 00:32:27,720 3/10 of a second. But you're also going into a 544 00:32:27,728 --> 00:32:32,029 regime that there is no data. I have to actually get this data 545 00:32:32,029 --> 00:32:35,527 from simulations, do some wind tunnel to validate that it's 546 00:32:35,535 --> 00:32:40,224 there. So that, so whether that's 547 00:32:40,232 --> 00:32:45,416 viable, you know, they have $150 million a year budget. 548 00:32:45,424 --> 00:32:48,500 I don't know, whatever the cap is, it's a little bit higher 549 00:32:48,508 --> 00:32:50,722 than that. Now on that budget, there's no 550 00:32:50,730 --> 00:32:54,610 way you could do this, right? Just collect just money there to 551 00:32:54,610 --> 00:32:59,743 collect the data you'd need. So, so it's, it's, it's a really 552 00:32:59,743 --> 00:33:02,445 interesting question. I, I, there's a group of people 553 00:33:02,445 --> 00:33:05,274 that fundamentally feel I will just spend so much money to 554 00:33:05,282 --> 00:33:09,500 collect the data so that I can make it just like language. 555 00:33:09,508 --> 00:33:12,944 But physics isn't I physics isn't language either. 556 00:33:12,952 --> 00:33:17,461 So if it, so there's still a bit here that we have to figure out. 557 00:33:17,461 --> 00:33:20,000 So my own personal belief is, or 558 00:33:20,000 --> 00:33:25,240 at least what I'm going to do is embed physics type knowledge 559 00:33:25,480 --> 00:33:27,760 into these neural networks so we can do this much more 560 00:33:27,760 --> 00:33:34,760 efficiently anyway. Yeah, no, I I've had this, you 561 00:33:34,760 --> 00:33:37,480 know, debate many time and you know, that was part of the 562 00:33:37,480 --> 00:33:40,619 reason for doing that paper with Johannes and Sid was to try 563 00:33:40,619 --> 00:33:44,880 and like try and project out some of this economies of 564 00:33:44,880 --> 00:33:47,602 things. But we still, I think even 565 00:33:47,610 --> 00:33:50,520 writing that paper, we, we still got to the end. 566 00:33:50,528 --> 00:33:53,358 And you know, I can't, I can't speak to the others, but at 567 00:33:53,366 --> 00:33:55,800 least for myself, we still didn't really resolve the 568 00:33:55,808 --> 00:34:00,440 fundamental challenge, which is even with all the predictions we 569 00:34:00,440 --> 00:34:05,864 made, they weren't reaching the accuracy that was 570 00:34:05,864 --> 00:34:12,030 necessarily required. And the some of that accuracy is 571 00:34:12,030 --> 00:34:16,397 very deep within the domain knowledge of those companies. 572 00:34:16,405 --> 00:34:21,469 And so a Formula One team would argue that they are the only 573 00:34:21,478 --> 00:34:25,965 ones who deeply know how to, you know, set up a a case and get it 574 00:34:25,965 --> 00:34:28,967 to give the right accuracy. If you just brute force even 575 00:34:28,976 --> 00:34:34,060 take an LES model and try and run it even with something like 576 00:34:34,060 --> 00:34:38,800 a billion cells, you could quite easily get it very wrong, 577 00:34:38,800 --> 00:34:42,440 very wrong. And even if you gave it 10 578 00:34:42,447 --> 00:34:45,222 billion doesn't actually, but you know you'd you almost then 579 00:34:45,230 --> 00:34:49,496 have to extrapolate it all the way to the end, which is to do a 580 00:34:49,496 --> 00:34:52,130 complete DNS or something. But then what if you don't know 581 00:34:52,130 --> 00:34:55,936 the porosity values of the radiator that they're using? 582 00:34:55,944 --> 00:35:03,103 So this I kind of feel that in one sense I agree with you that 583 00:35:03,103 --> 00:35:06,354 this does feel a little bit like just a data issue and then, 584 00:35:06,354 --> 00:35:09,520 you know, a bit like ChatGPT and others to show that, you 585 00:35:09,520 --> 00:35:13,568 know, just put a lot of money behind something you can solve 586 00:35:13,568 --> 00:35:17,640 it. But that I guess the difference 587 00:35:17,640 --> 00:35:20,400 I would say, which is why I think I would agree more with 588 00:35:20,400 --> 00:35:25,440 your approach is it makes sense when the opportunity of the 589 00:35:25,440 --> 00:35:28,320 market is trillions of dollars, which is basically enterprise 590 00:35:28,320 --> 00:35:34,480 AI, but CFD, I don't think it's a trillion dollar market. 591 00:35:34,480 --> 00:35:39,000 It's probably more than the 10s of billions maybe. 592 00:35:39,000 --> 00:35:44,120 So if you spend 10s of billions that makes you assume that you 593 00:35:44,120 --> 00:35:48,000 can take the entire market, which you know so from a like 594 00:35:48,000 --> 00:35:55,280 return on investment that to me is a harder like balance. 595 00:35:55,520 --> 00:35:59,920 Well, and, and it's also interesting for like for some of 596 00:35:59,920 --> 00:36:01,280 these companies, we need to do science. 597 00:36:01,280 --> 00:36:04,080 And I think you hit on such a critical issue, which is this 598 00:36:04,080 --> 00:36:08,080 idea of tolerances. There's you're always going to 599 00:36:08,080 --> 00:36:12,440 have to have, I think, recourse to a physical manifestation of 600 00:36:12,440 --> 00:36:16,440 what you're doing. Because if you're if OK. 601 00:36:16,720 --> 00:36:21,000 So think about world models a little bit, right? 602 00:36:21,680 --> 00:36:23,960 They're just fakes. They look cool. 603 00:36:25,760 --> 00:36:29,760 But you know, people who build world models, no lives are 604 00:36:29,760 --> 00:36:32,160 dependent upon it. It's not like by going here I 605 00:36:32,160 --> 00:36:36,920 could I could die, But if you have these people develop in the 606 00:36:36,920 --> 00:36:39,280 world model, I want you to design an airplane for me. 607 00:36:39,360 --> 00:36:41,360 Like there's no way I'm getting on that airplane. 608 00:36:43,320 --> 00:36:46,000 So there's there's there's sort of these zero failure 609 00:36:46,000 --> 00:36:50,440 environments where the tight tolerance is absolutely king. 610 00:36:50,560 --> 00:36:52,920 And even I think Bezos is quite interested in this. 611 00:36:53,080 --> 00:36:55,640 You know, again, I'm just throwing out this project 612 00:36:55,640 --> 00:36:58,120 Prometheus, because I'm just the, the staggering amount of 613 00:36:58,120 --> 00:37:01,480 money that's been invested in it in the space. 614 00:37:02,920 --> 00:37:05,840 But you know, he's interested in space travel and rockets and 615 00:37:05,840 --> 00:37:11,040 you're like, OK, so those are 0 failure environments. 616 00:37:11,920 --> 00:37:19,320 And I, you know, when, when a prompt goes bad like on ChatGPT 617 00:37:19,320 --> 00:37:23,000 or Gemini or Claude, nobody gets hurt, right? 618 00:37:23,000 --> 00:37:24,240 Really. I mean, at the end of the day, 619 00:37:24,240 --> 00:37:27,720 it's like, Oh, it didn't quite get the I didn't die cause of 620 00:37:27,720 --> 00:37:30,640 it. But like when you're going to do 621 00:37:30,640 --> 00:37:33,560 something like a rocket, right, where people's lives are at 622 00:37:33,560 --> 00:37:36,520 stake and you have to have the tolerances or else it's going to 623 00:37:38,040 --> 00:37:42,440 that's where I think there's just such a an amazing pressure, 624 00:37:42,760 --> 00:37:48,720 right, to get this right. And and it's not clear to me how 625 00:37:48,720 --> 00:37:52,320 AI does that, right. It could be that agentic systems 626 00:37:52,320 --> 00:37:54,600 will learn how to do like that's. 627 00:37:54,680 --> 00:37:56,960 I think we can certainly program. 628 00:37:56,960 --> 00:38:00,160 I think agentic systems to say now that I got most the answer 629 00:38:00,160 --> 00:38:06,840 right, my agents are all about doing the tolerance checking and 630 00:38:06,840 --> 00:38:09,240 figuring out what experiments have to be run. 631 00:38:10,360 --> 00:38:12,800 Because in my view, I think really what happens is I think 632 00:38:12,800 --> 00:38:15,720 the agentic system should come back to you and say, hey, Neil, 633 00:38:16,400 --> 00:38:18,720 we need to do some experiments because right now I'm starting 634 00:38:18,720 --> 00:38:21,840 to hallucinate or at least I'm uncertain my models. 635 00:38:23,000 --> 00:38:25,200 And of course it they're not going to actually talk to you. 636 00:38:25,200 --> 00:38:27,720 They're going to actually talk to their robot friends and say, 637 00:38:27,720 --> 00:38:31,600 can you do this experiment? Yeah, The kind of result, you 638 00:38:31,600 --> 00:38:34,160 know, I need you to do it under these kind of load conditions. 639 00:38:34,160 --> 00:38:36,960 So they can, once I get that back, I can proceed along with 640 00:38:36,960 --> 00:38:41,680 my iteration. So I think, I think that's going 641 00:38:41,680 --> 00:38:45,520 to be key for us somewhere along the way, right. 642 00:38:45,520 --> 00:38:49,080 But this tolerance idea is we still don't have a proof of, and 643 00:38:49,080 --> 00:38:51,040 it's not good enough. Just have a world model that 644 00:38:51,040 --> 00:38:57,320 looks cool and looks right. Yeah, that's, yeah. 645 00:38:57,320 --> 00:39:01,520 I think that's fundamentally the also I guess the practical 646 00:39:01,520 --> 00:39:06,000 challenge of this where I feel probably individual companies 647 00:39:06,000 --> 00:39:10,200 will, will still play more of a role in this. 648 00:39:11,000 --> 00:39:14,400 You know, if you're a large aircraft manufacturer, I would 649 00:39:14,400 --> 00:39:18,680 imagine that at least in the, you know, short to medium term, 650 00:39:18,680 --> 00:39:22,720 you're going to be the one that is most likely using some of 651 00:39:22,720 --> 00:39:27,040 your data to be fine-tuned this or, or you'll be still do it. 652 00:39:27,040 --> 00:39:30,160 You know, it's, it's, I don't think you completely outsource 653 00:39:30,160 --> 00:39:39,080 this to, to A, to a separate, you know, entity that that's 654 00:39:39,200 --> 00:39:41,360 feels, although you raise interesting point of 655 00:39:41,360 --> 00:39:43,840 experimental data because I think that's probably often 656 00:39:43,840 --> 00:39:48,640 overlooked that the, I guess the analogy, all the simulation is 657 00:39:48,640 --> 00:39:51,877 just synthetic data and it's not actually the ground truth, 658 00:39:51,877 --> 00:39:55,600 which is why I always struggle with initiatives to, you know, 659 00:39:55,600 --> 00:39:58,408 if I train a machine learning model, we've had this debate, 660 00:39:58,408 --> 00:40:00,015 I'd be interested to know your thought. 661 00:40:00,023 --> 00:40:03,399 So if we take like a standard data set and the goal of the 662 00:40:03,407 --> 00:40:06,470 exercise is that your surrogate model should be able to predict 663 00:40:06,470 --> 00:40:09,448 the ground truth and whoever gets it the closest gets the 664 00:40:09,456 --> 00:40:12,440 highest score and wins. But that. 665 00:40:12,448 --> 00:40:17,657 Has a certain ground truth baked into it because it was done with 666 00:40:17,657 --> 00:40:21,092 a RANS model and LES model. So if someone comes along who's 667 00:40:21,092 --> 00:40:24,720 trained this big foundation and then they try and predict your 668 00:40:24,728 --> 00:40:28,120 case, they could actually have a worse score because their method 669 00:40:28,120 --> 00:40:30,665 was trained on the different underlying simulation data. 670 00:40:30,673 --> 00:40:33,864 But then which one's right? Because you've never actually 671 00:40:33,872 --> 00:40:37,140 said what's the ground truth? The like the real world truth, 672 00:40:37,148 --> 00:40:38,455 you know? Yeah. 673 00:40:38,463 --> 00:40:43,856 Well, actually I I think this is also one of the very difficult 674 00:40:43,864 --> 00:40:47,680 and fundamental challenges in science and engineering We are 675 00:40:47,688 --> 00:40:51,242 right now using the ground truth is my simulation often like when 676 00:40:51,242 --> 00:40:53,900 we test our methods, you know, including us, you know, it's 677 00:40:53,908 --> 00:40:56,890 like, how will this method work? Well, I'll pretend the truth is 678 00:40:56,890 --> 00:41:01,363 my simulator. The problem I see is that we 679 00:41:01,371 --> 00:41:04,744 say, OK, well, what if you had the experiment? 680 00:41:04,752 --> 00:41:09,142 Well, in experiments we actually can never also have access to 681 00:41:09,150 --> 00:41:13,020 the ground truth because in an experiment you have sensors or 682 00:41:13,028 --> 00:41:18,160 either point sensors or you're measuring some observables, but 683 00:41:18,160 --> 00:41:24,069 you never have the full state information like you do in the 684 00:41:24,077 --> 00:41:27,496 simulation. So this idea of a ground truth 685 00:41:27,504 --> 00:41:31,869 is really interesting because you never have it in real 686 00:41:31,877 --> 00:41:35,105 systems. You have you have some 687 00:41:35,113 --> 00:41:38,580 manifestations of of measurements of that ground 688 00:41:38,588 --> 00:41:42,900 truth through your sensors, your full state knowledge is only 689 00:41:42,908 --> 00:41:44,772 through the in a simulation world. 690 00:41:44,780 --> 00:41:48,690 And how this connection of how do you assimilate those two 691 00:41:48,698 --> 00:41:50,549 together becomes now very important. 692 00:41:50,557 --> 00:41:54,900 And so somehow this, you know, the whole data simulation effort 693 00:41:54,900 --> 00:41:57,695 and we've seen the success of it and whether right. 694 00:41:57,703 --> 00:42:02,448 So a lot of companies have gotten into it because I think 695 00:42:02,456 --> 00:42:05,832 they, they benefited from basically the idea that like 696 00:42:05,840 --> 00:42:08,840 we've been collecting weather model weather data for like 697 00:42:08,848 --> 00:42:11,155 decades upon decades, very well resolved. 698 00:42:11,163 --> 00:42:15,720 And so they, they can train these big models and a lot of 699 00:42:15,728 --> 00:42:18,210 the people working that space like look what we can do with 700 00:42:18,218 --> 00:42:20,980 machine learning. It's like you understand you're 701 00:42:20,980 --> 00:42:25,840 in a very not representative scientific field, but for 40 702 00:42:25,848 --> 00:42:29,932 years, 50 years, we've been collecting data ad nauseam in 703 00:42:29,940 --> 00:42:34,205 this space and that's why your models work whereas and most 704 00:42:34,213 --> 00:42:37,176 other scientific fields, it's just not there. 705 00:42:37,184 --> 00:42:42,814 Or as you've pointed out, a lot of a lot of people guard their 706 00:42:42,822 --> 00:42:45,080 data closely. Like if you're for Formula One 707 00:42:45,088 --> 00:42:47,220 team, you don't share your simulation data. 708 00:42:47,228 --> 00:42:52,280 If you're a car company, you typically don't share your 709 00:42:52,288 --> 00:42:55,205 simulation data. People don't share their supply 710 00:42:55,205 --> 00:42:58,982 chain data like there. There's just so much data that 711 00:42:58,990 --> 00:43:02,474 is there, but nobody shares it. So we have that. 712 00:43:02,482 --> 00:43:04,972 It's kind of off limits a little bit. 713 00:43:04,980 --> 00:43:08,770 Which is quite I've often joked that probably the best company 714 00:43:08,778 --> 00:43:11,768 to make a foundation model for automotive is an automotive 715 00:43:11,776 --> 00:43:14,768 company. You know, frankly, if I was a 716 00:43:14,776 --> 00:43:18,300 car company, I'd be half tempted if I, you know, if I wasn't 717 00:43:18,308 --> 00:43:20,184 worried more about the car situation at the moment. 718 00:43:20,192 --> 00:43:22,927 I guess that's not their priority to build a foundation 719 00:43:22,935 --> 00:43:25,230 model. But like, you know, that's the 720 00:43:25,238 --> 00:43:27,318 ironic thing. They've got all the expertise, 721 00:43:27,326 --> 00:43:31,360 all the data, all the knowledge, all the facilities and they 722 00:43:31,368 --> 00:43:35,120 literally have it sat there, wind tunnel probably running 723 00:43:35,128 --> 00:43:37,022 24/7. Every all the, the stuff that, 724 00:43:37,030 --> 00:43:41,327 you know, a startup would dream of having, they actually have 725 00:43:41,335 --> 00:43:44,118 there. So it's yeah on that. 726 00:43:44,126 --> 00:43:46,960 What about your time at McLaren then? 727 00:43:46,968 --> 00:43:49,682 What? What I guess you must be into 728 00:43:49,690 --> 00:43:53,080 cars to, to also, you know, want to do that. 729 00:43:53,088 --> 00:43:57,240 But what did it teach you going from, you know, being a faculty 730 00:43:57,240 --> 00:43:59,332 member to taking that time at McLaren? 731 00:43:59,340 --> 00:44:04,912 How did it shape your thinking? Yeah, so, so first I was, it was 732 00:44:04,912 --> 00:44:08,792 such a privilege to be there. I I've been a Formula One fan 733 00:44:08,800 --> 00:44:11,436 since I was a little kid, so I'm half Brazilian. 734 00:44:11,444 --> 00:44:15,015 So in the 1970s, growing up a kid in Brazil, Emerson 735 00:44:15,023 --> 00:44:17,368 Fittipaldi was a Formula One champion. 736 00:44:17,376 --> 00:44:19,709 So every Brazilian loved Fittipaldi. 737 00:44:19,717 --> 00:44:24,075 But then I started watching in high school, in the first 738 00:44:24,083 --> 00:44:29,392 Formula One race I watched was a Ayrton Senna, and I was just 739 00:44:29,392 --> 00:44:32,640 some Sunday morning early. I started watching this race. 740 00:44:32,648 --> 00:44:36,554 It was on ESPN and there was this young Brazilian kid. 741 00:44:36,562 --> 00:44:41,195 Ayrton Senna won that race and I just became a super fan of 742 00:44:41,203 --> 00:44:42,986 Senna. And, and you know, of course 743 00:44:42,994 --> 00:44:45,710 then I'd loved McLaren because he was a three time champion 744 00:44:45,718 --> 00:44:47,874 with McLaren. And so I, I, I was a Formula One 745 00:44:47,874 --> 00:44:51,644 fan from the early days. And so I was on a sabbatical 746 00:44:51,652 --> 00:44:55,280 here in London and, and I thought, you know, what, if I 747 00:44:55,288 --> 00:44:58,175 can connect up there. And the awesome thing is a 748 00:44:58,183 --> 00:45:01,428 Spencer Sherwin who is at Imperial College in aerospace 749 00:45:01,436 --> 00:45:06,750 engineering, one of his former PhD students is head of CFD, 750 00:45:06,758 --> 00:45:10,645 Julien Hoessler. So it's just he connected me and 751 00:45:10,645 --> 00:45:15,416 then I was was out there and it taught me two things. 752 00:45:15,416 --> 00:45:19,936 Two really big take home messages that happened for me at 753 00:45:19,936 --> 00:45:23,148 McLaren that actually shaped part of why I even came to 754 00:45:23,148 --> 00:45:25,360 London to come to Autodesk number one. 755 00:45:26,280 --> 00:45:29,560 And this was really memorable. And it's, it's almost sad to say 756 00:45:29,560 --> 00:45:34,160 a little bit, but they were a team, they worked together. 757 00:45:34,240 --> 00:45:37,040 They had a common objective. You could talk to anybody from 758 00:45:37,040 --> 00:45:40,680 the person making coffee to the cafeteria people to head of 759 00:45:40,680 --> 00:45:45,117 engineering to the the brand new engineer, everybody knew and 760 00:45:45,117 --> 00:45:51,640 had a shared common goal. And we're really working 761 00:45:51,640 --> 00:45:55,520 together to that goal. And it's interesting coming from 762 00:45:55,520 --> 00:46:00,760 the academic environment, which is, yes, we're collaborative, 763 00:46:00,800 --> 00:46:06,557 but really we are ultra selfish in our, in our structure, 764 00:46:06,557 --> 00:46:12,790 right? And I, I love that I loved being 765 00:46:12,790 --> 00:46:18,698 around these people who wanted to do great work, who had a 766 00:46:18,706 --> 00:46:21,380 shared vision. That doesn't mean people didn't 767 00:46:21,380 --> 00:46:24,368 have, I'm not saying that everything was, you know, 768 00:46:24,376 --> 00:46:28,700 perfect, but I'm but that was felt to me so much healthier 769 00:46:28,708 --> 00:46:31,295 than what I've been in for decades. 770 00:46:31,303 --> 00:46:36,785 So that was the so that that that really stuck with me. 771 00:46:36,793 --> 00:46:40,888 I have to say it was a really, it was a really and, and just to 772 00:46:40,888 --> 00:46:44,150 for my own self reflection about like, Oh, I didn't even 773 00:46:44,150 --> 00:46:50,280 realize how much of A world I live in of self centeredness. 774 00:46:50,288 --> 00:46:52,895 So. That is the negative part of 775 00:46:52,903 --> 00:46:56,335 academia, isn't it? I mean, you know that sort of I 776 00:46:56,335 --> 00:46:59,420 always, I guess the best description I give, it feels 777 00:46:59,428 --> 00:47:02,600 like every professor is running their own startup or their own, 778 00:47:02,600 --> 00:47:05,332 you know, because it's very like people often call it 779 00:47:05,332 --> 00:47:09,280 something lab or some, you know, it's, it's and I guess at 780 00:47:09,280 --> 00:47:13,792 worse, you know, good luck being a head of department or a 781 00:47:13,792 --> 00:47:16,760 faculty. It's like all rivals going after 782 00:47:16,760 --> 00:47:18,320 each other, challenging for funding. 783 00:47:18,720 --> 00:47:21,600 I guess in the best scenario, it's an amazing cross 784 00:47:21,600 --> 00:47:25,440 collaborative environment of of of of people. 785 00:47:26,600 --> 00:47:28,240 But yeah, I, I see what you mean. 786 00:47:28,240 --> 00:47:31,320 There's a certain. I think I was very collaborative 787 00:47:31,320 --> 00:47:34,160 and I was very friendly across and worked with lots of people. 788 00:47:34,160 --> 00:47:38,000 But it's still was this very much like I still have to get my 789 00:47:38,000 --> 00:47:39,160 grant. Yeah. 790 00:47:39,360 --> 00:47:42,680 And I still have to take care of my students and, and this was 791 00:47:42,680 --> 00:47:46,080 just kind of refreshing, but there was this piece there. 792 00:47:46,080 --> 00:47:48,640 And then the thing that I came there, I was like, oh, you know, 793 00:47:48,720 --> 00:47:50,520 I've been doing reduced order models. 794 00:47:50,520 --> 00:47:52,880 I'm super excited about this to do things. 795 00:47:54,000 --> 00:48:00,240 But really the big goal there was to update shape of the car. 796 00:48:00,240 --> 00:48:02,200 And then you just ask a very simple question. 797 00:48:02,200 --> 00:48:06,680 Here's a, a mesh, which is this car, which is, let's call it a 798 00:48:06,680 --> 00:48:10,099 billion parameter mesh. How do you how do you do 799 00:48:10,099 --> 00:48:12,760 gradient descent updates on a three-dimensional mesh? 800 00:48:13,200 --> 00:48:16,840 And then I just saw that what you had there was a group of 801 00:48:16,840 --> 00:48:18,680 experts. And so it was human gradient 802 00:48:18,680 --> 00:48:24,760 descent, like true knowledge base of people who had long time 803 00:48:24,760 --> 00:48:28,320 experience understanding the flow physics off the front wing, 804 00:48:28,320 --> 00:48:32,040 the back wing, the underside, the side pods who were making 805 00:48:32,040 --> 00:48:37,080 collectively a decision about like what's the next upgrade of 806 00:48:37,080 --> 00:48:42,080 the shape that we try and even understanding when they felt 807 00:48:42,080 --> 00:48:44,680 like we have to go to the wind tunnel with this design, the 808 00:48:44,680 --> 00:48:47,320 wind tunnel is very expensive. We get very limited time. 809 00:48:48,280 --> 00:48:52,840 But this recourse to reality that would sort of pin them 810 00:48:53,440 --> 00:48:56,480 their designs down or actually invalidate them, right? 811 00:48:57,080 --> 00:49:01,560 It was really interesting to see this and I felt honestly, I 812 00:49:01,560 --> 00:49:05,840 guess I felt super like, I feel like I come in, I got this, you 813 00:49:05,840 --> 00:49:09,520 know, tool set and I felt powerless a little bit, which is 814 00:49:09,800 --> 00:49:15,720 here is this inverse design problem can't help with and and 815 00:49:15,720 --> 00:49:17,680 it really stuck with me afterwards. 816 00:49:17,680 --> 00:49:21,360 Like, how do I get into this geometry regime and understand 817 00:49:21,360 --> 00:49:23,840 how to handle geometry in a much better way? 818 00:49:23,840 --> 00:49:26,200 And that, you know, how do I get a latent representation of 819 00:49:26,200 --> 00:49:30,720 geometry to do updates? And so, you know, when Autodesk 820 00:49:30,720 --> 00:49:33,600 came along, it was like, wait a minute, this, this is the most, 821 00:49:33,760 --> 00:49:37,024 this is probably one of the best geometry companies in the 822 00:49:37,024 --> 00:49:41,720 world and do this design. And so it felt a very natural 823 00:49:41,720 --> 00:49:45,360 thing to come over because it was intellectually for me was 824 00:49:45,360 --> 00:49:48,120 one of the main things I wanted to go after. 825 00:49:48,880 --> 00:49:51,880 And I felt like if I try to do it in the academic environment, 826 00:49:51,880 --> 00:49:55,400 this pivot towards trying to generate all the skill set 827 00:49:55,400 --> 00:49:59,400 around geometry would take me a decade, right, to get like 828 00:49:59,720 --> 00:50:03,040 really there. Whereas coming to Autodesk was 829 00:50:03,040 --> 00:50:06,120 like, I immediately came into a group of colleagues who were 830 00:50:06,640 --> 00:50:09,160 they, you know, live and breathe this stuff. 831 00:50:09,160 --> 00:50:12,640 So it's like, amazing. So those are the two things that 832 00:50:12,640 --> 00:50:19,102 McLaren that really stuck out to me and and also maybe the 833 00:50:19,102 --> 00:50:23,120 third was just to walk into this McLaren tech center where. 834 00:50:23,840 --> 00:50:26,720 Quite nice. Very nice and also just to see 835 00:50:26,720 --> 00:50:29,960 like, you know, we rarely as an academics also see that, you 836 00:50:29,960 --> 00:50:32,080 know, by the time we write a paper, it's a year till it gets 837 00:50:32,080 --> 00:50:36,120 public. Like there's there's no here on 838 00:50:36,120 --> 00:50:39,320 this factory floor. What every, all these engineers 839 00:50:39,320 --> 00:50:43,170 could see was I've helped build this thing and it's right 840 00:50:43,170 --> 00:50:44,869 there. It's like this physical 841 00:50:44,877 --> 00:50:48,090 manifestation of beauty. And it's like a gallery, an art 842 00:50:48,090 --> 00:50:49,980 gallery, right? You're like, Oh my gosh, this is 843 00:50:49,980 --> 00:50:53,516 the. And even to see Senna's car, his 844 00:50:53,516 --> 00:50:57,592 MP4/4 from when he, I think his first championship, it's 845 00:50:57,600 --> 00:51:02,280 like it's, it's so motivating, right, To have that creative 846 00:51:02,288 --> 00:51:04,020 inspiration that's there. Yeah. 847 00:51:04,028 --> 00:51:08,256 Yeah, I know. It's I, I find that was my 848 00:51:08,264 --> 00:51:12,027 experience coming into Formula One and I've seen almost exactly 849 00:51:12,027 --> 00:51:14,899 the same thing. And I, I think almost every 850 00:51:14,907 --> 00:51:19,099 start up that comes in trying to sell into an F1 team faces the 851 00:51:19,107 --> 00:51:22,599 same reality that they all think, oh, surely, you know, 852 00:51:22,607 --> 00:51:27,306 you, you just need to do this or, or you know, like you need 853 00:51:27,314 --> 00:51:30,900 to use this better turbulence model or you need to do this 854 00:51:30,900 --> 00:51:33,362 surrogate model. And yeah, you realize that This 855 00:51:33,362 --> 00:51:36,837 is why I always say about foundation models or something 856 00:51:36,845 --> 00:51:41,520 that the idea that you come in and you would use a method to 857 00:51:41,528 --> 00:51:44,195 come up with a brand new design from scratch. 858 00:51:44,203 --> 00:51:49,109 In reality, they already know what needs to be done and it's 859 00:51:49,117 --> 00:51:54,500 tweaking the most fine things that is often very is never 860 00:51:54,508 --> 00:51:57,308 usually tested. It's usually like a, a big 861 00:51:57,316 --> 00:51:59,362 thing, isn't it? Like the delta between this and 862 00:51:59,362 --> 00:52:03,001 this and you're looking for the difference of, you know, 2 drag 863 00:52:03,001 --> 00:52:07,080 counts or something. It's, yeah, I've often found it 864 00:52:07,080 --> 00:52:10,180 quite humbling when you go in there thinking you can make a 865 00:52:10,188 --> 00:52:12,978 difference and then you're like, oh, this is already pretty well 866 00:52:12,978 --> 00:52:16,203 optimized. And also to see these people who 867 00:52:16,203 --> 00:52:19,664 have put in the time and effort and have the passion and 868 00:52:19,664 --> 00:52:23,682 they have that special skill set, like I know how to squeeze 869 00:52:23,682 --> 00:52:26,680 out a delta here in these manipulations. 870 00:52:26,688 --> 00:52:32,352 That sort of are the the the way I almost think about it is if we 871 00:52:32,352 --> 00:52:35,112 think about what AI is largely trained on, RMSE. 872 00:52:35,112 --> 00:52:37,800 RMSE is sort of a blunt instrument. 873 00:52:37,800 --> 00:52:41,800 It's sort of this, it's sort of I got all the big scale stuff 874 00:52:41,800 --> 00:52:44,480 for you, right. But the winning race car is 875 00:52:44,480 --> 00:52:49,680 about all the fine details and promoting these in the loss 876 00:52:49,680 --> 00:52:52,280 function. Somehow, you know, this thinks 877 00:52:52,280 --> 00:52:56,040 I, you know, you can see that neural networks act act as 878 00:52:56,040 --> 00:52:59,160 bandpass filters essentially. Like I got the big stuff right. 879 00:52:59,160 --> 00:53:02,080 All the small stuff doesn't really contribute to that RMSE 880 00:53:02,080 --> 00:53:04,800 score. So like, whatever, we'll throw 881 00:53:04,800 --> 00:53:06,640 it out. Now, of course we can make 882 00:53:06,640 --> 00:53:09,240 efforts to, you know, put a diffusion model down there or 883 00:53:09,240 --> 00:53:12,000 something like this to try to like fill in in so it makes it 884 00:53:12,000 --> 00:53:17,960 look right. But the this attention to small 885 00:53:17,960 --> 00:53:23,920 detail which a human can make the focus, maybe we teach our 886 00:53:23,920 --> 00:53:28,102 neural networks to do this, but like in their general 887 00:53:28,102 --> 00:53:32,336 deployment that's is not really, you know, these little micro 888 00:53:32,336 --> 00:53:36,288 adjustments are so fine detail, they don't score in a training 889 00:53:36,288 --> 00:53:39,440 score hardly. Now that is where they'll like 890 00:53:39,440 --> 00:53:44,920 do think the agents I am far more bullish on because I feel 891 00:53:44,920 --> 00:53:49,720 like with the greatest respect to the Formula One engineers, 892 00:53:49,760 --> 00:53:53,760 what they're doing is, is still something that is just you can 893 00:53:53,760 --> 00:53:57,800 describe what they do and and actually, I think if you look 894 00:53:57,800 --> 00:54:03,480 across the grid, my hypothesis of why is one team, you know, 895 00:54:03,480 --> 00:54:06,680 one year amazing and then another year it's not, you know, 896 00:54:06,680 --> 00:54:09,840 like if they really understood everything that's needed to be 897 00:54:09,840 --> 00:54:13,160 done, surely they would just each year build too great. 898 00:54:13,400 --> 00:54:16,680 The fact that they take wrong turns, they go down the wrong 899 00:54:16,680 --> 00:54:19,680 direction. They have a a very large 900 00:54:19,680 --> 00:54:22,760 optimization space that once you start going down a route, you 901 00:54:22,760 --> 00:54:25,880 can't really turn back and you sort of have to go. 902 00:54:25,880 --> 00:54:30,976 And I do feel that it is still a very large optimization 903 00:54:30,976 --> 00:54:34,100 problem that a machine could ultimately do better than a 904 00:54:34,100 --> 00:54:37,800 human in in theory with ultimate resources. 905 00:54:37,808 --> 00:54:40,216 But. Yeah, yeah. 906 00:54:40,224 --> 00:54:44,598 So if if I was, if I was in somehow in charge of some 907 00:54:44,606 --> 00:54:48,221 version of some formula that one team and sort of as a let's say 908 00:54:48,221 --> 00:54:51,880 call it iOS, I'm the AI director of of some form on 909 00:54:51,880 --> 00:54:55,080 team. What I think I would be doing, 910 00:54:55,080 --> 00:54:59,000 and I'm also bullish on the agent piece of this. 911 00:54:59,000 --> 00:55:03,400 I would be setting up extensive interviews with all of my 912 00:55:03,400 --> 00:55:06,880 experienced engineers. Like how are they making their 913 00:55:06,880 --> 00:55:08,800 decision? So you made this decision. 914 00:55:08,800 --> 00:55:11,719 Why I want you to verbalize it. 915 00:55:11,727 --> 00:55:15,809 Why did you make decision? What do you see in the data? 916 00:55:15,817 --> 00:55:19,269 Because a lot of the experiential learning that 917 00:55:19,277 --> 00:55:27,274 these, you know, master engineers have, they have a 918 00:55:27,282 --> 00:55:30,155 pipeline. It's just they haven't maybe 919 00:55:30,163 --> 00:55:32,660 verbalized it. And if you could pull that out 920 00:55:32,668 --> 00:55:36,262 and you say my agent's going to try to take on your persona the 921 00:55:36,262 --> 00:55:39,440 way you think, but it needs to be trained in the way you think, 922 00:55:39,440 --> 00:55:42,056 the way you're making your decision points, I actually 923 00:55:42,064 --> 00:55:44,987 think that is a viable way forward. 924 00:55:44,995 --> 00:55:48,754 And that could accelerate things significantly. 925 00:55:48,762 --> 00:55:55,440 And so the downloading of experience of people, somehow 926 00:55:55,448 --> 00:55:59,694 some next step that needs to happen in it. 927 00:55:59,702 --> 00:56:02,950 And it could be that these people have never thought about 928 00:56:02,950 --> 00:56:06,664 like deeply about what was my algorithm if I had to, if I have 929 00:56:06,664 --> 00:56:09,655 to write down what I do and when I'm looking at the car. 930 00:56:09,663 --> 00:56:11,575 And because it's actually interesting, when you look at 931 00:56:11,583 --> 00:56:13,635 McLaren, you can walk up and down this pit of engineers. 932 00:56:13,643 --> 00:56:17,040 And then when you're down on the design aero side, you see 933 00:56:17,048 --> 00:56:20,122 basically people have the double screen like they do at every 934 00:56:20,122 --> 00:56:23,094 tech company and what they're got pictures on there is flow 935 00:56:23,102 --> 00:56:25,299 physics. And these guys are just studying 936 00:56:25,299 --> 00:56:28,275 this all day long to make their decisions. 937 00:56:28,283 --> 00:56:34,096 But if you could understand, if they could verbalize what 938 00:56:34,104 --> 00:56:37,582 they're doing in that study, you know what machine could do it 939 00:56:37,590 --> 00:56:40,028 Like, and just like, OK, I did in five seconds. 940 00:56:40,036 --> 00:56:44,068 What you all morning looking at here at 5 seconds later. 941 00:56:44,076 --> 00:56:46,260 Here's here's here's the assessment. 942 00:56:46,268 --> 00:56:48,342 Yes. If you don't know what they're 943 00:56:48,350 --> 00:56:51,892 looking at, it's really hard for the agent to to know that, 944 00:56:51,900 --> 00:56:54,440 right. The only thing on that and then, 945 00:56:54,440 --> 00:56:58,535 and I want to go too down a rabbit hole on this one, but is 946 00:56:58,535 --> 00:57:01,860 if I was an experienced F1 engineer who's done it for 20 947 00:57:01,868 --> 00:57:05,328 years, I'd almost want to, you know, copyright my skills.md 948 00:57:05,336 --> 00:57:09,940 or something or, or monetize it. Because, you know, as you could 949 00:57:09,940 --> 00:57:12,271 imagine, like once that person's described F1 process 950 00:57:12,271 --> 00:57:15,584 and you put it into some skills file or whatever, and 951 00:57:15,584 --> 00:57:17,940 then the team goes, right, well, thank you very much. 952 00:57:17,948 --> 00:57:20,544 The agent actually now runs much faster than you can. 953 00:57:20,552 --> 00:57:24,626 So, you know, see you later. I do often want, because I 954 00:57:24,634 --> 00:57:27,988 completely agree with you, this human knowledge can be 955 00:57:27,996 --> 00:57:31,536 extracted, you know, because it is often quite repeatable. 956 00:57:31,544 --> 00:57:34,428 But I, I think it's an interesting, you know, if 957 00:57:34,436 --> 00:57:38,350 someone's creating a startup of like, you know, I, I'm going to 958 00:57:38,350 --> 00:57:43,240 monetize my knowledge and yeah, use me as an agent, but I need 959 00:57:43,248 --> 00:57:46,606 some money out of it because. Yeah. 960 00:57:46,614 --> 00:57:49,902 So this is actually, I think this, this argument is one of 961 00:57:49,910 --> 00:57:54,109 the most amazing, I think things we're going to have to deal with 962 00:57:54,109 --> 00:57:56,840 overall in the tech community. Even, you know, here at 963 00:57:56,848 --> 00:57:58,828 Autodesk, they're in America, which is OK. 964 00:57:58,836 --> 00:58:02,935 So if these agents are so effective and we can program 965 00:58:02,943 --> 00:58:07,384 them to be effective, if you're a Formula One team, do you say 966 00:58:07,392 --> 00:58:11,965 like, if I can do this, I could shrink my workforce or do you go 967 00:58:11,965 --> 00:58:16,424 the other ways with my workforce, I can 10X the number 968 00:58:16,424 --> 00:58:21,925 of designs I'm doing. Going through like this is a 969 00:58:21,933 --> 00:58:26,859 really interesting point for us. Like a company like here we say 970 00:58:26,859 --> 00:58:30,025 like, well, hey, you know, essentially one person can do 971 00:58:30,033 --> 00:58:32,940 now 10X the work they used to be able to. 972 00:58:32,948 --> 00:58:36,560 Well, guess what, we could take 10X customers. 973 00:58:36,568 --> 00:58:40,166 Yeah, yeah. Or we could cut the company down 974 00:58:40,166 --> 00:58:41,840 by 10. Yeah. 975 00:58:41,848 --> 00:58:45,282 So. I, I suspect it will probably be 976 00:58:45,282 --> 00:58:49,442 more that you'll just be able to do more, you know, with the, 977 00:58:49,442 --> 00:58:51,756 with the, the staff that you have. 978 00:58:51,764 --> 00:58:57,096 I I, I suspect so that's my gut feeling because whenever we've 979 00:58:57,104 --> 00:59:02,652 had more, you know, your HPC facility can now run your cases 980 00:59:02,652 --> 00:59:05,126 twice as fast or the code can run. 981 00:59:05,134 --> 00:59:07,525 But normally it's been that, OK, great. 982 00:59:07,533 --> 00:59:10,942 Now you just have to do double the amount of work or you do 983 00:59:10,950 --> 00:59:13,200 twice as many simulations or you do whatever. 984 00:59:13,208 --> 00:59:17,195 But I, yeah, I definitely think that is a, an interesting one. 985 00:59:17,203 --> 00:59:20,987 The question for you on Autodesk, do you have any 986 00:59:20,995 --> 00:59:23,198 regrets? You didn't go into industry 987 00:59:23,206 --> 00:59:27,576 earlier on, like now that you're there, is there any? 988 00:59:27,584 --> 00:59:31,169 Yeah. Yeah, actually not so much. 989 00:59:31,177 --> 00:59:36,814 I mean, I, I, I actually, I think this is just a perfect 990 00:59:36,822 --> 00:59:41,852 time in life for me to come here at this point. 991 00:59:41,860 --> 00:59:51,430 But I, I, I think that, I mean, I guess in some sense I haven't 992 00:59:51,430 --> 00:59:54,640 been so money motivated overall, career wise. 993 00:59:54,648 --> 00:59:59,468 I think obviously could have gone into a tech company in 994 00:59:59,476 --> 01:00:02,103 Seattle much earlier. Like you pick one, they're all 995 01:00:02,111 --> 01:00:06,328 there. There wasn't I, I, I was more 996 01:00:06,336 --> 01:00:10,050 interested in doing really interesting work if I could. 997 01:00:10,058 --> 01:00:16,260 Autodesk kind of like I said, was a confluence of the right 998 01:00:16,268 --> 01:00:19,097 time intellectually. Like I said, I wanted to really 999 01:00:19,097 --> 01:00:22,200 think about how can I handle geometry and physics jointly? 1000 01:00:22,208 --> 01:00:24,790 How does geometry induce physics, right? 1001 01:00:24,798 --> 01:00:28,840 I was just so fascinated after that McLaren year that. 1002 01:00:28,848 --> 01:00:36,502 So it was, it was the right time intellectually for me to come to 1003 01:00:36,502 --> 01:00:41,500 to Autodesk and I'm very happy with it now. 1004 01:00:41,508 --> 01:00:44,360 But I don't regret my time at UW. 1005 01:00:44,368 --> 01:00:47,168 I think that was found, you know, what I learned there and 1006 01:00:47,176 --> 01:00:51,088 the students I had, I loved my students, my postdocs, my 1007 01:00:51,096 --> 01:00:53,894 collaborators. It's just, it's just a good time 1008 01:00:53,894 --> 01:00:55,634 in life. And also my kids graduated from 1009 01:00:55,634 --> 01:00:59,748 high school, so I didn't have to try to live in a good school 1010 01:00:59,748 --> 01:01:02,822 district in London. Me and my wife could just say, 1011 01:01:02,830 --> 01:01:06,165 where do we want to live? I went off to worry about like 1012 01:01:06,173 --> 01:01:08,990 all this kids stuff, right? Which you know, which is, you 1013 01:01:08,998 --> 01:01:10,334 know, I went through that already. 1014 01:01:10,342 --> 01:01:13,823 I'm done. So. 1015 01:01:13,831 --> 01:01:16,547 So where's the future lie then? Where? 1016 01:01:16,555 --> 01:01:19,361 Where do you see if we could put a looking glass? 1017 01:01:19,369 --> 01:01:23,564 And we skipped forward in five years and you and I talk again. 1018 01:01:23,564 --> 01:01:24,640 Hopefully it's somewhere like 1019 01:01:24,640 --> 01:01:27,200 Barcelona or some nice, you know, some nice location. 1020 01:01:28,240 --> 01:01:30,040 Well, where? Where do you think we'll be? 1021 01:01:32,040 --> 01:01:35,000 Yeah. So the first thing that I think 1022 01:01:35,000 --> 01:01:39,760 is going to happen sooner than later is people aren't going to 1023 01:01:39,760 --> 01:01:41,920 put out GitHub code. They're going to put out GitHub 1024 01:01:41,920 --> 01:01:44,880 agents. I'm just using that language. 1025 01:01:44,880 --> 01:01:47,560 Like, you know, right now you say like, hey, I wrote this code 1026 01:01:47,560 --> 01:01:49,680 you can download on GitHub. It's like, no, no, you're just 1027 01:01:49,680 --> 01:01:54,400 going to give me access to your agent that manages all of that, 1028 01:01:54,920 --> 01:01:56,480 right? So it feels to me like this 1029 01:01:56,480 --> 01:01:58,800 agentic push is quite a real thing. 1030 01:01:58,960 --> 01:02:01,840 Like even programming on a higher level, like with a 1031 01:02:02,160 --> 01:02:04,400 something like Cline, just in terms of the workflow. 1032 01:02:04,400 --> 01:02:08,800 Instead of me starting to just run your code, I just, I'm 1033 01:02:08,800 --> 01:02:10,520 interacting with your agent, with your code. 1034 01:02:11,840 --> 01:02:14,720 I think that's one thing that is bound to happen. 1035 01:02:14,720 --> 01:02:17,960 I, I, maybe I'm wrong, but I, I kind of feel like that's right. 1036 01:02:17,960 --> 01:02:22,080 That's just the such a clean pathway for people to share code 1037 01:02:22,080 --> 01:02:25,080 as you're sharing the agent. So it's not like it's just like 1038 01:02:25,080 --> 01:02:27,960 you're saying, here's my code and the grad student that wrote 1039 01:02:27,960 --> 01:02:29,560 it on. So if you have any questions, 1040 01:02:29,560 --> 01:02:31,840 they can answer all of it because they built this code. 1041 01:02:32,240 --> 01:02:36,120 But it's like agentic part, the agentic part of that. 1042 01:02:36,120 --> 01:02:43,040 No Second, I, I just feel like we're going to have like this 1043 01:02:43,160 --> 01:02:48,880 ability to deploy so many of these agents and partnerships 1044 01:02:48,880 --> 01:02:50,640 that we're going to be able to come up with. 1045 01:02:51,120 --> 01:02:55,520 I, I think we're going to have much bigger thoughts than we've 1046 01:02:55,520 --> 01:02:59,640 had in the past. Because in a lot of our future 1047 01:02:59,640 --> 01:03:03,240 thinking, we don't just think, what if we did this? 1048 01:03:03,480 --> 01:03:06,280 We also have to balance it. Like, yeah, but what could I 1049 01:03:06,280 --> 01:03:08,902 actually maybe do if I stretch? 1050 01:03:08,910 --> 01:03:15,040 But now what you could do with you stretch is it's it's so much 1051 01:03:15,040 --> 01:03:20,320 bigger. So I think our I think about our 1052 01:03:20,320 --> 01:03:24,397 goal setting capabilities now in terms of where do. 1053 01:03:24,405 --> 01:03:28,800 And I think I'm still trying to get my head around that now 1054 01:03:28,808 --> 01:03:32,340 because I'm trying to also train myself to think much bigger 1055 01:03:32,348 --> 01:03:35,969 thoughts about what we could achieve given the tools that 1056 01:03:35,977 --> 01:03:39,624 have just just even in one year have emerged. 1057 01:03:39,632 --> 01:03:44,470 Like one year alone has all of a sudden give you gives you this 1058 01:03:44,478 --> 01:03:49,552 transformational ability. And the hard part about making 1059 01:03:49,560 --> 01:03:54,146 these projections into the future is that right now, you 1060 01:03:54,154 --> 01:03:58,680 don't know, like if someone's going to all of a sudden pop 1061 01:03:58,688 --> 01:04:01,410 something out and like one month, by the way, here's this 1062 01:04:01,418 --> 01:04:04,920 new tool, check it out. And it's like, Oh my gosh, which 1063 01:04:04,920 --> 01:04:07,928 is everything, right? Like this is starting to happen 1064 01:04:07,928 --> 01:04:11,260 even in design space. So I'll promote a paper that 1065 01:04:11,268 --> 01:04:17,120 these these guys wrote from sort of a largely Oxford-Cambridge 1066 01:04:17,128 --> 01:04:19,240 collaboration called Articraft. 1067 01:04:19,248 --> 01:04:23,165 Maybe you saw this one. This is just this generative 1068 01:04:23,173 --> 01:04:26,845 design with articulated engineering products and it's 1069 01:04:26,853 --> 01:04:31,248 just like this fascinating thing that they were able to build out 1070 01:04:31,248 --> 01:04:35,268 and you're like, OK, I didn't think this was this was a cry. 1071 01:04:35,276 --> 01:04:37,772 Like incredible that they achieved it, right. 1072 01:04:37,780 --> 01:04:43,565 And it just feels like, So what people are able to achieve 1073 01:04:43,573 --> 01:04:49,374 sometimes is beyond what I like, like I'm not imagining big 1074 01:04:49,382 --> 01:04:50,556 enough. Frankly, it is. 1075 01:04:50,564 --> 01:04:53,938 I think that's my the the take home message for myself in this 1076 01:04:53,938 --> 01:04:58,165 last year is like I've got to be much more grand scoped in my 1077 01:04:58,165 --> 01:05:00,200 imagination about what could be. 1078 01:05:02,480 --> 01:05:05,240 And it's also part of what I'm stealing with our team is like, 1079 01:05:05,280 --> 01:05:10,600 we need to think a lot bigger. Yes, some things I I would I 1080 01:05:10,600 --> 01:05:14,040 mean that's been probably one of the things that I've enjoyed 1081 01:05:14,040 --> 01:05:17,480 being NVIDIA is it's a company that, you know, obviously thanks 1082 01:05:17,480 --> 01:05:20,720 to the guy the top Jensen, you know, tends to think quite big 1083 01:05:20,720 --> 01:05:25,080 and it is quite infectious. You know, you do start to now, 1084 01:05:25,080 --> 01:05:27,280 of course, that's going to be based on delivery, you know, 1085 01:05:27,320 --> 01:05:29,760 that you can deliver it and it has to have some reality. 1086 01:05:30,040 --> 01:05:35,440 But I think, yeah, thinking big is probably something that in 1087 01:05:35,440 --> 01:05:37,240 some ways academia is good at doing. 1088 01:05:37,240 --> 01:05:41,440 But I feel like to your earlier point is also one where we're 1089 01:05:41,440 --> 01:05:44,386 quite quick to shut things down, you know, especially the 1090 01:05:44,386 --> 01:05:48,692 review process and the sort of there is a little bit of a 1091 01:05:48,692 --> 01:05:53,280 skepticism or so, you know, whereas probably the tech 1092 01:05:53,280 --> 01:05:57,993 world is more willing to like, which is I guess why all 1093 01:05:57,993 --> 01:06:01,596 startups come about. And, and yeah, so I I would, I 1094 01:06:01,604 --> 01:06:05,652 would tend to agree with you that thinking big is actually a 1095 01:06:05,652 --> 01:06:09,460 requirement at the moment given how fast things are moving. 1096 01:06:09,468 --> 01:06:12,876 Yeah. And and also I, I, I think that 1097 01:06:12,876 --> 01:06:17,107 academics used to have some of the bigger thought life, I guess 1098 01:06:17,107 --> 01:06:21,202 I would say. And there's still some truth to 1099 01:06:21,202 --> 01:06:25,188 that, but at least in our fields, it's not clear they have 1100 01:06:25,188 --> 01:06:27,512 the resources now. Well, that's. 1101 01:06:27,520 --> 01:06:32,940 To, to go after, and this is partly the success, what I've 1102 01:06:32,948 --> 01:06:37,280 seen of the computer science crowd is if you really look at 1103 01:06:37,288 --> 01:06:41,356 some of the big pushes and some of the academics involved, it's 1104 01:06:41,356 --> 01:06:44,920 because they've been in partnerships with resource rich 1105 01:06:44,920 --> 01:06:49,170 companies like Google, like Meta, like NVIDIA, where it's 1106 01:06:49,178 --> 01:06:53,062 like, like if they were just sitting at Stanford, like I have 1107 01:06:53,062 --> 01:06:55,824 my friends there, Ali's lab. One of my grad students is going 1108 01:06:55,824 --> 01:06:57,952 to go there for a postdoc and I'm super excited. 1109 01:06:57,960 --> 01:07:01,152 I was like, you know, it's like, if you want to do something 1110 01:07:01,160 --> 01:07:05,437 interesting there, it's pretty easy to reach out and I think 1111 01:07:05,445 --> 01:07:10,060 have these people partnering with you to do big scale work. 1112 01:07:10,068 --> 01:07:14,528 Like if you were just there like me and you have to write a grant 1113 01:07:14,528 --> 01:07:18,162 to get a little machine that like can only do a fraction of 1114 01:07:18,170 --> 01:07:22,535 it, like, but now they can they can really work with in this 1115 01:07:22,543 --> 01:07:29,346 environment to do great things. So yeah, I, I think right now 1116 01:07:29,354 --> 01:07:33,252 industry is favored in terms of transformational parts. 1117 01:07:33,260 --> 01:07:40,568 And because the resources are there to do things, I think the 1118 01:07:40,568 --> 01:07:46,410 pressure on industry is to get the right partnerships with 1119 01:07:46,418 --> 01:07:49,335 academic people, right and vice versa. 1120 01:07:49,343 --> 01:07:53,502 Like so if you I think that still a really great strategy is 1121 01:07:53,502 --> 01:07:57,652 that, you know, even here at Autodesk, I have definitely am 1122 01:07:57,660 --> 01:08:01,698 reaching out to people that I think are really valuable and 1123 01:08:01,706 --> 01:08:06,386 making connections. So I'm valuable to them now 1124 01:08:06,394 --> 01:08:10,080 sitting on the industry side, they're valuable to me sitting 1125 01:08:10,088 --> 01:08:12,540 on sort of the intellectual thought life side. 1126 01:08:12,548 --> 01:08:16,611 But the partnership is fantastic because everybody really wins in 1127 01:08:16,611 --> 01:08:19,697 it, yes. And so that's, that's I think a 1128 01:08:19,697 --> 01:08:21,951 a really important place to go forward to. 1129 01:08:21,959 --> 01:08:25,359 Yeah, No, I I agree with you that the I should probably 1130 01:08:25,368 --> 01:08:26,812 should rephrase what I said before. 1131 01:08:26,819 --> 01:08:29,800 It's true academia can have big thoughts, but because they know 1132 01:08:29,800 --> 01:08:32,332 they don't have the resources, it's almost like, well, that's a 1133 01:08:32,332 --> 01:08:34,836 nice forward, but I'm never going to be able to do it. 1134 01:08:34,844 --> 01:08:38,148 So what can I actually achieve and get a paper out and, you 1135 01:08:38,156 --> 01:08:41,448 know, get this funding in? So they then have to think 1136 01:08:41,456 --> 01:08:44,281 smaller, whereas you're right, if you're a, you know, a big 1137 01:08:44,288 --> 01:08:46,194 tech company, you can be a bit bolder. 1138 01:08:46,203 --> 01:08:51,062 So the when the two come together, you get the best, I 1139 01:08:51,069 --> 01:08:53,462 guess. So maybe as a final question for 1140 01:08:53,462 --> 01:08:58,192 you, and we touched upon it a little bit, if you're, we have, 1141 01:08:58,192 --> 01:09:02,276 let's say, some students listening now, whether they're 1142 01:09:02,283 --> 01:09:06,765 undergraduates or PhDs, it's a tricky time, right? 1143 01:09:06,773 --> 01:09:12,198 You know, what would you recommend that they would focus, 1144 01:09:12,198 --> 01:09:16,710 let's say, a PhD on or what should they study to be relevant 1145 01:09:16,710 --> 01:09:21,040 in the next sort of five years of this wave of transformation? 1146 01:09:21,040 --> 01:09:24,120 Yeah, so my first thing I tell 1147 01:09:24,120 --> 01:09:31,200 them is they're living, I think in one of the most exciting 1148 01:09:31,840 --> 01:09:38,147 times in human history because I think this this time, the 1149 01:09:38,147 --> 01:09:41,072 future generations, they'll pinpoint this period of time 1150 01:09:41,072 --> 01:09:45,399 is like the world changed. And this is massively 1151 01:09:45,399 --> 01:09:49,680 influential about what what it means for us as humanity right 1152 01:09:49,680 --> 01:09:51,720 now. This is and to be part of it. 1153 01:09:51,920 --> 01:09:53,560 I mean, it doesn't mean it has a good ending. 1154 01:09:53,640 --> 01:09:55,440 Whatever. I'm I'm just saying that you but 1155 01:09:55,440 --> 01:09:58,040 they are part of the they are part of the puzzle piece in 1156 01:09:58,040 --> 01:10:01,480 there. And not only so it's, it's 1157 01:10:01,480 --> 01:10:03,080 fascinating, It's a little scary. 1158 01:10:03,080 --> 01:10:05,000 They got a strap on their seat belt and go. 1159 01:10:05,040 --> 01:10:09,480 And just like this is going and there's no stopping this thing. 1160 01:10:09,480 --> 01:10:12,280 I mean, as much as people want to step back and say, let's 1161 01:10:12,280 --> 01:10:16,000 wait, let's talk about it, it's like it's too much inertia. 1162 01:10:16,000 --> 01:10:17,960 Is it? We're we're going and we just 1163 01:10:17,960 --> 01:10:22,800 have to do this. But there's this famous quote of 1164 01:10:23,000 --> 01:10:26,520 from Picasso that I always like to share with people. 1165 01:10:27,880 --> 01:10:31,800 It's what was from 1968, the year I was born, and Picasso 1166 01:10:31,800 --> 01:10:33,960 1968. He made a comment about 1167 01:10:33,960 --> 01:10:37,640 computers and he his his quote, computers are worthless. 1168 01:10:38,280 --> 01:10:40,900 They can only answer questions. 1169 01:10:40,908 --> 01:10:47,217 And I think that statement is amazing today, which is the real 1170 01:10:47,217 --> 01:10:52,840 value, I think is us as humans are still there asking the what 1171 01:10:52,840 --> 01:10:55,375 ifs. You know, a lot of these 1172 01:10:55,383 --> 01:10:57,880 startups come from people like, what if we could do this? 1173 01:10:57,888 --> 01:11:01,325 What if we could do that? What if I want to go to the 1174 01:11:01,333 --> 01:11:02,841 moon? I want to build a airplane, I 1175 01:11:02,849 --> 01:11:07,772 want to build a Formula One car? How would I So the largely it 1176 01:11:07,780 --> 01:11:12,714 feels like even with this AI kick, these are amazing new 1177 01:11:12,722 --> 01:11:16,200 tools. But really you're still in 1178 01:11:16,208 --> 01:11:20,200 charge of really directing where a lot of this goes. 1179 01:11:20,208 --> 01:11:22,410 You're not a passenger, you're the driver. 1180 01:11:22,418 --> 01:11:26,235 It's just that you now have a Lamborghini or a Ferrari, right? 1181 01:11:26,235 --> 01:11:28,800 And you didn't get taught how to 1182 01:11:28,800 --> 01:11:30,720 drive this thing. So that's what makes it a little 1183 01:11:30,720 --> 01:11:35,120 scared. So, so that's one thing the, 1184 01:11:35,200 --> 01:11:38,120 the, the other is that I still think you need foundational 1185 01:11:38,120 --> 01:11:39,720 knowledge. You still need to be as well 1186 01:11:39,720 --> 01:11:42,680 educated as you can and, you know, just really foundational 1187 01:11:42,680 --> 01:11:45,440 thinking, whether that's it's for the mathematics that 1188 01:11:45,440 --> 01:11:48,800 underlies, whether it's computational math, you know, 1189 01:11:48,840 --> 01:11:52,840 deep linear algebra knowledge or deep physics knowledge. 1190 01:11:52,840 --> 01:11:56,280 These things still matter a great deal because at some point 1191 01:11:56,280 --> 01:11:59,320 they're going to play fundamental roles in largely 1192 01:11:59,320 --> 01:12:03,640 developing your thought processes and your critical 1193 01:12:03,640 --> 01:12:05,720 thinking ability. But also maybe you bring you 1194 01:12:05,720 --> 01:12:11,000 back to foundational thinking it when you're developing these 1195 01:12:11,080 --> 01:12:13,800 these models. And of course for them, they 1196 01:12:13,800 --> 01:12:16,960 just have to adopt these tools like Claude. 1197 01:12:16,960 --> 01:12:19,680 If they're not using Claude Code or something like Cline or 1198 01:12:19,680 --> 01:12:24,240 Cursor, it's like every day you're getting further behind 1199 01:12:24,240 --> 01:12:26,880 from where people are working, right? 1200 01:12:28,560 --> 01:12:31,160 And it feels a little uncomfortable, right? 1201 01:12:31,320 --> 01:12:33,200 Ultimately for some of them, right? 1202 01:12:33,200 --> 01:12:35,480 It's like, I mean, I'm getting this code. 1203 01:12:36,040 --> 01:12:37,960 I don't should I read through it? 1204 01:12:37,960 --> 01:12:41,880 Should I read me this thousand lines of code and it works, but 1205 01:12:41,880 --> 01:12:44,920 I don't I kind of just only have a vague idea of what these 1206 01:12:44,920 --> 01:12:49,200 pieces doing that feels very uncomfortable because like I'm 1207 01:12:49,200 --> 01:12:51,355 sure when you were in school and when I was in school, it's 1208 01:12:51,355 --> 01:12:53,800 like you had to know every single line in your code. 1209 01:12:53,880 --> 01:12:58,200 You wrote every single your code as those such a transformation 1210 01:12:58,200 --> 01:13:01,600 of like, yes, you have to kind of know that you have to know 1211 01:13:01,600 --> 01:13:03,720 what's going on, but on the other hand, you just like to 1212 01:13:03,720 --> 01:13:08,400 figure out how to move at the speed of what's happening. 1213 01:13:08,400 --> 01:13:11,920 So like and it's such a 2 opposite poles that have to be 1214 01:13:11,920 --> 01:13:13,880 there. I did. 1215 01:13:15,040 --> 01:13:18,160 I was having a conversation with one professor who mentioned 1216 01:13:18,160 --> 01:13:19,920 something interesting. I won't say his name just in 1217 01:13:19,920 --> 01:13:22,640 case he didn't want it to be shared, but although it's 1218 01:13:22,640 --> 01:13:27,440 nothing controversial, it was just a point of how would you, 1219 01:13:28,400 --> 01:13:32,840 if you're an undergraduate, say you should use AI as a tutor, 1220 01:13:33,440 --> 01:13:36,560 not to do it for you. So you should still do 1221 01:13:36,560 --> 01:13:39,600 everything yourself, but essentially use AI to guide you. 1222 01:13:39,600 --> 01:13:43,019 It's like they are unbelievably good at explaining things in 1223 01:13:43,019 --> 01:13:46,528 any tone you want in any adapter, you know, like, but 1224 01:13:46,528 --> 01:13:49,222 you've got to do it yourself because if not, you won't 1225 01:13:49,222 --> 01:13:52,146 learn it. And then as you transition into, 1226 01:13:52,146 --> 01:13:57,322 let's say, a PhD, it starts to become more of a, an assistant, 1227 01:13:57,322 --> 01:14:01,550 you know, where maybe you have learnt some of the physics. 1228 01:14:01,558 --> 01:14:05,686 So now you can rely on it more and you're just sort of checking 1229 01:14:05,686 --> 01:14:07,630 it. And then as you get even more 1230 01:14:07,638 --> 01:14:11,434 late in your career, you're almost using it as a PhD or as a 1231 01:14:11,434 --> 01:14:14,392 junior engineer. And so I feel like the risk is 1232 01:14:14,400 --> 01:14:17,080 if you use AI at the undergraduate level like 1233 01:14:17,088 --> 01:14:21,348 somebody does, maybe in our situation that's the danger 1234 01:14:21,356 --> 01:14:24,740 because then you've not really learnt it. 1235 01:14:24,748 --> 01:14:29,192 But of course the temptation is to use it when you're an 1236 01:14:29,200 --> 01:14:31,565 undergraduate in that way. And that's probably the risk 1237 01:14:31,573 --> 01:14:34,080 factor, isn't it? Like, do you ever really 1238 01:14:34,088 --> 01:14:36,925 understand things when you have this cheat code? 1239 01:14:36,933 --> 01:14:40,196 Since you start playing computer game with the cheat code, you 1240 01:14:40,204 --> 01:14:42,800 know you it's hard not to use it sometimes. 1241 01:14:42,808 --> 01:14:47,446 Well, and, and I think the I think that's a fair assessment 1242 01:14:47,454 --> 01:14:50,360 of things. And I think ultimately if, if, 1243 01:14:50,368 --> 01:14:55,529 if you know, if you, if I were still an academic academia and 1244 01:14:55,537 --> 01:15:00,048 so forth, I think where I think the, the thought really needs to 1245 01:15:00,048 --> 01:15:03,946 be spent on how to use this tool is the high end students 1246 01:15:03,946 --> 01:15:06,460 that you're super smart students, you don't have to 1247 01:15:06,460 --> 01:15:08,440 worry about them. They'll just figure it all out. 1248 01:15:08,800 --> 01:15:10,400 They didn't need you in the 1st place. 1249 01:15:10,400 --> 01:15:12,840 They can just your stuff out. They'll learn that on their own. 1250 01:15:12,840 --> 01:15:16,080 They'll have deep knowledge of stuff and they'll, they're not a 1251 01:15:16,080 --> 01:15:18,600 concern. The low end students have always 1252 01:15:18,600 --> 01:15:21,320 been problematic because they never quite get it no matter how 1253 01:15:21,320 --> 01:15:24,040 much you try to, you know, they're maybe not spending the 1254 01:15:24,040 --> 01:15:26,640 time they need. It's that middle group who are 1255 01:15:26,640 --> 01:15:29,200 going to be your day-to-day engineers and so many companies. 1256 01:15:30,480 --> 01:15:34,653 How do you educate that group to be good stewards of the 1257 01:15:34,653 --> 01:15:36,776 software of the practices, right? 1258 01:15:36,784 --> 01:15:42,140 Because I think that group needs probably the most guidance of 1259 01:15:42,148 --> 01:15:47,983 how to be an intelligent engineer in a world where it is 1260 01:15:47,983 --> 01:15:52,000 so tempting to just pawn it off here, right? 1261 01:15:52,008 --> 01:15:56,386 It's easy and I get the right hand, you know, whatever, but it 1262 01:15:56,386 --> 01:16:01,220 but how, how do you teach that group to use these things 1263 01:16:01,228 --> 01:16:04,735 responsibly so that they're so they are, you know, because 1264 01:16:04,743 --> 01:16:07,845 look, these are the people who build our airplanes and build 1265 01:16:07,853 --> 01:16:12,704 our cars like it's we need, we need them to be proficient and 1266 01:16:12,712 --> 01:16:15,836 be responsible, right, because we are going to be in their 1267 01:16:15,844 --> 01:16:17,054 product space. Yeah, yeah. 1268 01:16:17,062 --> 01:16:20,848 And so, and I don't know what the quite the right answer there 1269 01:16:20,848 --> 01:16:23,124 is, except that we need them to know these tools. 1270 01:16:23,132 --> 01:16:26,904 We need them to also balance it with like, yeah, but you check 1271 01:16:26,912 --> 01:16:31,420 these tools and have maturity about using them because people 1272 01:16:31,420 --> 01:16:35,420 depend upon you. Check because I'm going to get 1273 01:16:35,428 --> 01:16:37,390 in the car you built me. And yeah. 1274 01:16:37,398 --> 01:16:39,980 And I don't want that thing falling apart when I'm, you 1275 01:16:39,988 --> 01:16:41,800 know, going down the freeway. Right. 1276 01:16:41,808 --> 01:16:44,492 Yeah, exactly. Great. 1277 01:16:44,500 --> 01:16:46,470 Well, thank you so much for taking the time to speak. 1278 01:16:46,478 --> 01:16:50,099 I I'm sure we could have carried on for many hours and I hope we 1279 01:16:50,099 --> 01:16:53,090 can do that. But maybe over, you know, a 1280 01:16:53,098 --> 01:16:54,140 drink sometime. Yeah. 1281 01:16:54,148 --> 01:16:58,355 That sounds good to me. But yeah, all the best. 1282 01:16:58,363 --> 01:17:00,442 It's Autodesk. I'm very excited to see what 1283 01:17:00,450 --> 01:17:03,742 your group's going to create and I hope, I'm sure we'll, we'll 1284 01:17:03,750 --> 01:17:05,820 hear about it over the coming years. 1285 01:17:05,828 --> 01:17:07,630 Awesome. Thank you again. 1286 01:17:07,638 --> 01:17:08,664 Great. You got it. 1287 01:17:08,672 --> 01:17:09,160 Thanks, Neil.