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 --> 00:00:30,040 the lessons they've learned on the way that I hope will be 10 00:00:30,040 --> 00:00:33,800 helpful to you too. So sit back and enjoy this 11 00:00:33,800 --> 00:00:41,160 episode. Hi and welcome back to the Neil 12 00:00:41,160 --> 00:00:43,840 Ashton Podcast. It's been a while, but I'm 13 00:00:43,840 --> 00:00:48,960 excited to start a new season and series and I'm going to 14 00:00:48,960 --> 00:00:54,480 start with a discussion on how I think agents and foundation 15 00:00:54,480 --> 00:00:59,440 models are are really a disruptor to the current CAE and 16 00:00:59,440 --> 00:01:03,040 EDA ecosystem. This, this is of course is just 17 00:01:03,040 --> 00:01:05,480 a personal opinion. This does not represent my 18 00:01:05,480 --> 00:01:08,960 current employer, NVIDIA, but I wanted to share some, some 19 00:01:08,960 --> 00:01:13,000 thoughts I've really had over the past six months of how I see 20 00:01:13,000 --> 00:01:15,916 this technology being, you know, obviously a massive 21 00:01:15,916 --> 00:01:18,420 opportunity, but also also a disruptor. 22 00:01:18,428 --> 00:01:22,220 And so we'll, we'll walk through that today from a technical 23 00:01:22,228 --> 00:01:25,681 point of view, from a commercial point of view, from a scientific 24 00:01:25,681 --> 00:01:28,026 point of view. And I'd certainly be interested 25 00:01:28,026 --> 00:01:31,464 to see whether you agree with my, my thinking on this. 26 00:01:31,472 --> 00:01:36,561 If we take a step back and you know, I'm going to look at as 27 00:01:36,569 --> 00:01:40,208 often I do in this, in this podcast through a sort of CFD 28 00:01:40,216 --> 00:01:44,429 lens, but I think much of what I'm saying is relevant across 29 00:01:44,437 --> 00:01:48,290 CAE and and also EDA for chip design, etcetera. 30 00:01:48,298 --> 00:01:51,543 If you look, where did things begin? 31 00:01:51,551 --> 00:01:56,795 I would argue in the 60s and the 70s people were writing their 32 00:01:56,803 --> 00:02:00,460 own codes. Now, I've had episodes with 33 00:02:00,468 --> 00:02:06,287 Professor Jameson and others who who started writing codes 34 00:02:06,295 --> 00:02:10,500 themselves before modern supercomputing was around, and 35 00:02:10,508 --> 00:02:13,912 those often started at academic institutions. 36 00:02:13,920 --> 00:02:18,015 Those codes then obviously started to become useful and 37 00:02:18,023 --> 00:02:21,374 there was collaborations with industry, particularly in the 38 00:02:21,382 --> 00:02:25,444 aerospace initially. And that is why that then those 39 00:02:25,444 --> 00:02:29,149 aerospace companies started to work closely with those 40 00:02:29,157 --> 00:02:32,890 developers and eventually started to take them as their 41 00:02:32,898 --> 00:02:35,810 own codes. And that's why you see today 42 00:02:35,818 --> 00:02:38,580 even there's a legacy of aerospace companies having their 43 00:02:38,580 --> 00:02:41,652 own codes written from scratch, frankly, because there 44 00:02:41,652 --> 00:02:45,540 were no other options. And that was the way to do it. 45 00:02:45,548 --> 00:02:49,754 The automotive sector then took some things from the aerospace 46 00:02:49,762 --> 00:02:51,496 sector. And this depends on this a 47 00:02:51,504 --> 00:02:53,264 little bit, whether we're talking about structural 48 00:02:53,272 --> 00:02:57,615 mechanics or fluid mechanics, but it started off being you had 49 00:02:57,615 --> 00:03:00,740 to write it yourself with some expert developer. 50 00:03:00,748 --> 00:03:06,280 Then over the course of, you know, the 70s, eighties and 51 00:03:06,288 --> 00:03:09,600 early 90s, those codes started to become commercialized. 52 00:03:09,608 --> 00:03:14,612 Those professors out of places like Imperial or MIT or Stanford 53 00:03:14,612 --> 00:03:17,640 started to, to, to create companies. 54 00:03:17,648 --> 00:03:22,772 And, and ultimately those are the, the lineage, the lineage of 55 00:03:22,772 --> 00:03:26,389 where Abaqus comes from, where Fluent comes from, where 56 00:03:26,397 --> 00:03:31,455 OpenFOAM comes from. And in some ways those 57 00:03:31,463 --> 00:03:35,970 commercialization was because it was a natural thing as one 58 00:03:35,978 --> 00:03:40,632 computing started to become and make these tools more than just 59 00:03:40,632 --> 00:03:44,380 a, a niche application. HPC facilities, you know, 60 00:03:44,388 --> 00:03:48,660 Beowulf clusters started to become commonplace and and of 61 00:03:48,668 --> 00:03:52,880 course, industry started to see the benefit of those simulation 62 00:03:52,880 --> 00:03:55,575 approaches. And frankly, then there was also 63 00:03:55,575 --> 00:03:59,032 a competitive nature that once one company starts, you start 64 00:03:59,040 --> 00:04:02,079 having competitions between, you know, the CD-adapcos and the, 65 00:04:02,087 --> 00:04:06,300 and what what is now the, the Ansys and the later the 66 00:04:06,308 --> 00:04:09,260 Dassaults. There was that growing ecosystem 67 00:04:09,260 --> 00:04:13,000 of companies that that compete against each other, that 68 00:04:13,008 --> 00:04:15,583 accelerates it. The computing starts to get 69 00:04:15,591 --> 00:04:19,678 better and other than those main players, you also started to 70 00:04:19,687 --> 00:04:21,800 have the hardware companies get involved. 71 00:04:21,808 --> 00:04:26,202 So the Intel's, the AMD, the Dell, the Lenovo, the Cray in 72 00:04:26,210 --> 00:04:30,167 those days, the the main people building those those systems, 73 00:04:30,175 --> 00:04:35,575 those supercomputing systems. And if you look at how it 74 00:04:35,583 --> 00:04:40,288 happened, that was probably true for, for maybe 5 to 10 years or 75 00:04:40,288 --> 00:04:44,710 so where it was mainly a creating software to run on 76 00:04:44,718 --> 00:04:46,654 workstations, servers and supercomputers. 77 00:04:46,662 --> 00:04:50,568 And that, you know, spread across the world, frankly. 78 00:04:50,576 --> 00:04:54,580 And those tools were very widely used and started to create a 79 00:04:54,588 --> 00:04:58,104 real, you know, hundreds of millions or even billions by 80 00:04:58,112 --> 00:05:02,660 then industry. The the next technology that 81 00:05:02,668 --> 00:05:07,165 really came around was then around like cloud computing. 82 00:05:07,173 --> 00:05:12,768 So cloud computing maybe 5-6, seven years ago was, was a 83 00:05:12,776 --> 00:05:16,160 disruptor because it allowed and, and obviously I'm skipping 84 00:05:16,160 --> 00:05:18,565 huge other things that happened, but I don't want to 85 00:05:18,565 --> 00:05:21,400 spend 2 hours just giving the history of, of CFD. 86 00:05:21,408 --> 00:05:26,762 The reason I pick out cloud, and I saw it first time working for 87 00:05:26,762 --> 00:05:31,354 AWS, is it changed it from being a purely software license 88 00:05:31,354 --> 00:05:36,380 point of view to where these providers could also have a 89 00:05:36,380 --> 00:05:39,970 control of the hardware. Packaging software and hardware 90 00:05:39,970 --> 00:05:44,460 together is attractive for a company for an ISV because they 91 00:05:44,460 --> 00:05:48,769 can make profit off the hardware and they can lock it 92 00:05:48,769 --> 00:05:51,146 together both from a performance point of view, but 93 00:05:51,146 --> 00:05:53,240 also from a sort of stickiness point of view. 94 00:05:53,920 --> 00:05:57,640 And there's advantages for the customer in terms of making it 95 00:05:57,640 --> 00:06:02,280 more seamless because often if you just ship a binary, they 96 00:06:02,280 --> 00:06:04,480 have to be the one to then figuring out how to get the 97 00:06:04,480 --> 00:06:06,480 hardware. They are the ones who've got to 98 00:06:06,640 --> 00:06:09,200 think about, do they need to buy twice as much hardware? 99 00:06:09,200 --> 00:06:12,920 If they want to run twice as many simulations in the theory 100 00:06:12,920 --> 00:06:17,160 of the ultimate SaaS, then it is fully scalable. 101 00:06:17,160 --> 00:06:21,124 So if you're running, you know, STAR-CCM+ or Fluent or 102 00:06:21,124 --> 00:06:26,200 PowerFLOW or whatever, you could essentially say, I want to run 103 00:06:26,200 --> 00:06:29,280 10 times more and through the, you know, infinite scale or 104 00:06:29,280 --> 00:06:31,560 inferior of the cloud you can do in the back end. 105 00:06:32,200 --> 00:06:37,760 So that SaaSification was, was something started maybe 6-7 106 00:06:37,760 --> 00:06:40,520 years ago. And, and obviously companies 107 00:06:40,520 --> 00:06:44,160 like Rescale were one of the first to do that at a cross 108 00:06:44,160 --> 00:06:46,389 software level. And there's others like 109 00:06:46,389 --> 00:06:49,960 TotalCAE, etcetera who tried to, to bring it together. 110 00:06:51,480 --> 00:06:56,400 That is where I would say AWS and Microsoft and Google started 111 00:06:56,400 --> 00:07:01,607 to enter into more seriously the CAE and EDA market as 112 00:07:01,607 --> 00:07:05,120 important partners. Because frankly, now it wasn't 113 00:07:05,120 --> 00:07:08,120 just The Dells and Lenovos, but you were deploying it on AWS or 114 00:07:08,120 --> 00:07:11,720 GCP or or Azure or or Oracle, for example. 115 00:07:12,320 --> 00:07:15,000 Now GPUs obviously had been around for quite a few years 116 00:07:15,000 --> 00:07:18,520 before that, but but arguably it's only been in the past sort 117 00:07:18,520 --> 00:07:23,440 of three, maybe four years where they have risen to become a 118 00:07:23,440 --> 00:07:25,560 really important part of the ecosystem. 119 00:07:26,160 --> 00:07:28,800 And frankly, that's because a lot of people have seen quite 120 00:07:28,800 --> 00:07:31,000 clear benefits and they're moving to the GPUs. 121 00:07:31,680 --> 00:07:35,280 And that's another technology point that's important because 122 00:07:36,000 --> 00:07:41,120 it added a new competitive angle to this because they needed to 123 00:07:41,200 --> 00:07:47,040 able to run things on a GPU, which meant that if you could 124 00:07:47,040 --> 00:07:49,400 optimize your code for GPU can get advantage. 125 00:07:49,400 --> 00:07:52,200 This is where Flexcompute, this is where Luminary Cloud started 126 00:07:52,200 --> 00:07:54,680 to come out. The rise of, you know, VC 127 00:07:54,680 --> 00:07:58,680 funding for these sort of companies led to a resurgence of 128 00:07:58,680 --> 00:08:01,240 competitiveness, I would say, in that market. 129 00:08:03,000 --> 00:08:05,840 And I the reason I point to the cloud before is that many of 130 00:08:05,840 --> 00:08:09,560 those companies lent on as the cloud being an additional sort 131 00:08:09,560 --> 00:08:12,040 of angle to their to their offering. 132 00:08:12,520 --> 00:08:19,480 So at that point another thing happens which is that the bigger 133 00:08:19,480 --> 00:08:24,800 companies start to look to consolidate and to start their 134 00:08:24,800 --> 00:08:28,040 sort of acquisition period. So you saw companies like 135 00:08:28,040 --> 00:08:31,520 Siemens and the timing of these are not all correct, but you 136 00:08:31,520 --> 00:08:37,679 know Siemens and Altair, Ansys and Synopsys, Cadence and BETA 137 00:08:37,679 --> 00:08:40,600 CAE and Pointwise, or Future Facilities. 138 00:08:41,440 --> 00:08:44,680 Many of these companies started to acquire each other. 139 00:08:45,720 --> 00:08:51,040 And why was that done? Well actually so those companies 140 00:08:51,600 --> 00:08:54,986 then wanted to start acquiring and of course this is where 141 00:08:54,986 --> 00:08:59,256 one of the key technology came in that starts to get to my 142 00:08:59,256 --> 00:09:02,424 point here, which is the AI side. 143 00:09:02,432 --> 00:09:07,348 It was, it has been around to be fair, for maybe six years or 144 00:09:07,356 --> 00:09:12,356 even 7 years. But really it was since the sort 145 00:09:12,356 --> 00:09:17,578 of ChatGPT moment in the early 2020s, 2021-2022 that brought 146 00:09:17,578 --> 00:09:22,322 this to life. And so, you know, they started 147 00:09:22,330 --> 00:09:28,189 some new startups to really get hold: Neural Concept, PhysicsX, 148 00:09:28,189 --> 00:09:34,528 NAVASTO, Extrality and Emmi AI. These companies started to 149 00:09:34,536 --> 00:09:37,774 come about offering the surrogate modelling capability, 150 00:09:37,774 --> 00:09:41,075 the ability to do real time simulations. 151 00:09:41,083 --> 00:09:45,740 And that frankly was a threat and an opportunity. 152 00:09:45,748 --> 00:09:50,020 And I say that because that is what I've heard when in 153 00:09:50,028 --> 00:09:52,052 conversations. It's a threat because anything 154 00:09:52,060 --> 00:09:54,360 new is by definition a potential threat. 155 00:09:54,368 --> 00:09:58,225 These software companies who have, you know, obviously had 156 00:09:58,233 --> 00:10:01,700 quite a, a good position, see as you know, how do we deal with 157 00:10:01,708 --> 00:10:03,728 this? But clearly it's also a massive 158 00:10:03,728 --> 00:10:06,400 opportunity because it brings new technology, new capabilities 159 00:10:06,400 --> 00:10:10,979 to, to, to customers. And that it has been a very 160 00:10:10,987 --> 00:10:14,664 interesting thing to observe in the past, let's say 3 or 4 years 161 00:10:14,664 --> 00:10:17,521 of how do these larger companies deal with it? 162 00:10:17,529 --> 00:10:21,904 How do companies deal with it who have their own codes? 163 00:10:21,912 --> 00:10:24,800 How do they respond? Do they respond with hype? 164 00:10:24,808 --> 00:10:27,140 Do they respond with negativity, with positivity? 165 00:10:27,148 --> 00:10:32,104 And what happened was then some companies decided to start 166 00:10:32,112 --> 00:10:34,970 developing themselves, some started to buy. 167 00:10:34,978 --> 00:10:38,501 So Ansys bought Extrality, which is now SimAI. 168 00:10:38,509 --> 00:10:42,570 Autodesk bought NAVASTO, Siemens, you could argue and 169 00:10:42,578 --> 00:10:45,430 some part of it was the Altair acquisition. 170 00:10:45,438 --> 00:10:50,944 And then some of these start off start-ups that started with a 171 00:10:50,952 --> 00:10:55,394 more GPU focus started to pivot quite heavily to the surrogate 172 00:10:55,402 --> 00:10:58,040 modelling. So Luminary Cloud or Luminary AI 173 00:10:58,040 --> 00:11:02,206 now and Flexcompute, both of you know pivoted to, to, to that 174 00:11:02,206 --> 00:11:06,615 direction. And so we, so we're at that 175 00:11:06,623 --> 00:11:11,500 point now where the, the ecosystem is at a very 176 00:11:11,508 --> 00:11:15,260 influential point for a sort of inflection point. 177 00:11:15,268 --> 00:11:19,032 And there's one final bit that is important to that. 178 00:11:19,040 --> 00:11:22,599 And it's actually ultimately the main topic that I want to speak 179 00:11:22,599 --> 00:11:26,900 about today, which is agents. Because the missing piece to 180 00:11:26,908 --> 00:11:32,340 some of this has been if you create a surrogate model and you 181 00:11:32,340 --> 00:11:36,432 take it to its fullest extent and you have some sort of 182 00:11:36,440 --> 00:11:41,027 foundation model, you haven't really got rid of the need to do 183 00:11:41,027 --> 00:11:43,937 CAD. You haven't got rid of the need 184 00:11:43,937 --> 00:11:46,672 to do some sort of preprocessing or post 185 00:11:46,672 --> 00:11:48,880 processing. You could argue you've just got 186 00:11:49,080 --> 00:11:54,355 a faster plug in replacement for the for the bit of the solver 187 00:11:54,355 --> 00:11:58,160 or the solve part, but you still need all the bit around the 188 00:11:58,160 --> 00:12:01,840 rest. And that's why it's ultimately 189 00:12:01,840 --> 00:12:06,200 been more of an integration play through some of these AI 190 00:12:06,200 --> 00:12:12,000 start-ups because you're not going to replace CFD or FEA or 191 00:12:12,040 --> 00:12:14,640 EDA, you're just really creating an additional tool. 192 00:12:15,320 --> 00:12:18,800 So in that sense you can understand why the larger ISVs 193 00:12:18,800 --> 00:12:22,760 and others have not necessarily seen this surrogate modelling 194 00:12:22,760 --> 00:12:26,520 as a massive threat initially because they know that on it's 195 00:12:26,520 --> 00:12:32,800 own it's it's not a replacement that has led to of course 1 196 00:12:32,800 --> 00:12:36,000 angle, which is if you were to make a model that was so 197 00:12:36,000 --> 00:12:41,360 generalizable, then it could given the right plug and play, 198 00:12:41,640 --> 00:12:44,360 disrupt some solver only companies. 199 00:12:45,000 --> 00:12:49,600 If you only produce a solver and not the CAD and maybe not the 200 00:12:49,600 --> 00:12:53,920 preprocessing, then arguably you could, you know, replace some 201 00:12:53,920 --> 00:12:58,720 of it. But the technology that I think 202 00:12:58,760 --> 00:13:05,480 it adds a real new dimension to this is the agentic side. 203 00:13:06,000 --> 00:13:09,240 Because what are what are agents? 204 00:13:09,240 --> 00:13:13,856 Well, first of all, let's tie it back to the usefulness of 205 00:13:13,856 --> 00:13:17,720 ChatGPT or Gemini. We have all found them to be 206 00:13:17,720 --> 00:13:22,760 incredibly useful in our day-to-day tasks, but I have 207 00:13:22,760 --> 00:13:27,080 personally noticed a a huge shift in usefulness with the 208 00:13:27,080 --> 00:13:28,960 agentic side. And what do we mean by the 209 00:13:28,960 --> 00:13:31,520 agentic side? At least the way I define it. 210 00:13:31,840 --> 00:13:36,600 Well, I think of it is just simply now the LLM is able to 211 00:13:36,600 --> 00:13:38,520 call some tools. Now that's probably 212 00:13:38,520 --> 00:13:42,200 oversimplifying it, but from a practical point of view, instead 213 00:13:42,200 --> 00:13:47,600 of me just being able to ask some ask Gemini to do something 214 00:13:47,600 --> 00:13:49,920 for me or, or Codex or whatever. 215 00:13:50,280 --> 00:13:53,080 Although the use of Codex now is sort of murky in the waters a 216 00:13:53,080 --> 00:13:54,680 little bit. If I stick to like a chat 217 00:13:54,680 --> 00:13:58,040 prompt, it's only going to be able to really do things that 218 00:13:58,040 --> 00:14:03,720 rely just on text or to A to a limit where what I really want 219 00:14:03,720 --> 00:14:06,600 it to do is to be able to go and do things for me. 220 00:14:07,080 --> 00:14:11,880 So for example, if I ask, I want you to create a summary of all 221 00:14:11,880 --> 00:14:13,200 my emails. Great. 222 00:14:13,480 --> 00:14:17,360 But what if I want you to now look through all my emails, look 223 00:14:17,360 --> 00:14:22,680 through all my Slack messages, and then I want you to go in and 224 00:14:24,040 --> 00:14:27,000 also look at a journal paper that I'm writing. 225 00:14:27,560 --> 00:14:31,680 And I want you to look at the emails where I got some response 226 00:14:31,680 --> 00:14:34,200 from somebody. And I want you to take that, go 227 00:14:34,200 --> 00:14:37,320 into, go into that. Oh, and I also want you to go 228 00:14:37,320 --> 00:14:39,840 and create some graphs for me. So can you also create some 229 00:14:39,840 --> 00:14:42,680 Python graphs using, you know, Matplotlib, etcetera. 230 00:14:43,240 --> 00:14:46,560 And I want you to do that all for me autonomously. 231 00:14:47,160 --> 00:14:51,480 What I'm describing to you now is absolutely achievable right 232 00:14:51,480 --> 00:14:53,880 this second using Claude Code or Codex. 233 00:14:55,200 --> 00:15:00,920 It is also possible for the tools that that that that LLM is 234 00:15:00,920 --> 00:15:04,280 essentially calling or the agent is calling can be simulation 235 00:15:04,280 --> 00:15:07,880 tools. So if I extend that further and 236 00:15:07,880 --> 00:15:11,280 I say I want you to be helping me to do some journal paper, 237 00:15:12,040 --> 00:15:16,120 maybe I could say now, well, could you actually, could you 238 00:15:16,120 --> 00:15:19,160 run some simulations for me? I have an idea. 239 00:15:19,320 --> 00:15:24,440 You know, here's some e-mail correspondence, here is a thread 240 00:15:24,480 --> 00:15:26,760 on a Slack channel with some collaborators. 241 00:15:27,880 --> 00:15:31,558 Not just write the paper, but will you go and set up an 242 00:15:31,558 --> 00:15:35,296 OpenFOAM simulation that that tries out some of those ideas 243 00:15:35,296 --> 00:15:37,240 we discussed and, and can you kick it off? 244 00:15:37,640 --> 00:15:39,609 And then when it's done, can you create the graphs and can you 245 00:15:39,609 --> 00:15:41,960 do it for me? That's completely achievable 246 00:15:41,960 --> 00:15:43,640 today. That is not future thinking. 247 00:15:43,640 --> 00:15:49,760 That is achievable today. My point is in what part of that 248 00:15:49,760 --> 00:15:57,120 then is this CFD tool? The main thing it's not is it 249 00:15:57,640 --> 00:16:02,240 because actually Codex or Claude or Vibe Code now with 250 00:16:02,240 --> 00:16:05,600 Mistral entering as a new sort of player in this field from 251 00:16:05,600 --> 00:16:09,000 with a heavy sort of sovereign AI angle, European angle. 252 00:16:09,440 --> 00:16:13,938 So Codex, Claude or Vibe, they can actually go and do that for 253 00:16:13,938 --> 00:16:16,000 you. And you don't have to be an 254 00:16:16,040 --> 00:16:21,840 expert user of OpenFOAM because most likely those models 255 00:16:21,840 --> 00:16:25,000 have either been trained on sufficient smart data that 256 00:16:25,000 --> 00:16:28,000 included OpenFOAM documentation and various 257 00:16:28,000 --> 00:16:32,440 reports across the web. Or more likely that in a real 258 00:16:32,440 --> 00:16:35,200 life scenario, and maybe this is where the commercial bit comes 259 00:16:35,200 --> 00:16:39,480 in, one of those providers could have specifically added that 260 00:16:39,480 --> 00:16:41,680 capability. They could have looked at 261 00:16:41,680 --> 00:16:46,185 science and engineering as a key area they want to get into, 262 00:16:46,185 --> 00:16:49,750 and so they will actually try to make their models good at 263 00:16:49,750 --> 00:16:53,840 doing this. At that point, the stickiness or 264 00:16:53,840 --> 00:16:58,825 the the difficulty of using a CFD code is no longer there, 265 00:16:58,825 --> 00:17:03,180 which is arguably one of the reasons those codes exist or 266 00:17:03,180 --> 00:17:06,880 have such a strong Moat because they're not easy to use. 267 00:17:07,160 --> 00:17:10,480 They require expertise, they're very well validated. 268 00:17:11,400 --> 00:17:15,280 If agent starts to come in and very powerful LLMs come in that 269 00:17:15,280 --> 00:17:19,839 can use these tools very well, then actually the uplift of 270 00:17:19,839 --> 00:17:23,040 changing to a different tool is not as much. 271 00:17:23,880 --> 00:17:28,280 And the ability to orchestrate multiple tools is also the same. 272 00:17:29,080 --> 00:17:34,280 So it raises a very interesting question, which is are the 273 00:17:34,640 --> 00:17:38,520 software companies just going to reside to being just tools and 274 00:17:38,520 --> 00:17:42,440 skills and actually your relationship will be more with 275 00:17:43,120 --> 00:17:47,640 the OpenAIs, the Mistrals, the Anthropics, etcetera. 276 00:17:48,400 --> 00:17:53,360 Or, and it is the case today, will new companies arise that 277 00:17:53,360 --> 00:17:55,840 are trying to do this? Interestingly, and there's very 278 00:17:55,840 --> 00:17:58,040 little public information on this, but you have companies 279 00:17:58,040 --> 00:18:01,859 like Project Prometheus, I think Jeff Bezos put out some 280 00:18:01,859 --> 00:18:05,152 announcement about wanting to create a artificial engineer 281 00:18:05,152 --> 00:18:07,954 out of this. So read into that what you will. 282 00:18:07,954 --> 00:18:09,280 But it sounds pretty clear that 283 00:18:09,280 --> 00:18:11,880 they want to build some sort of capability like that. 284 00:18:13,440 --> 00:18:17,178 You see the recent acquisition of Emmi AI by Mistral and you 285 00:18:17,178 --> 00:18:21,220 see a huge interest from OpenAI and Anthropic in general, AI 286 00:18:21,220 --> 00:18:24,868 for science, which makes you look at the future and start 287 00:18:24,868 --> 00:18:29,070 to think there could be a new wave now of companies coming 288 00:18:29,070 --> 00:18:32,490 around. And there are companies that are 289 00:18:32,490 --> 00:18:37,784 now base themselves purely to create agents to create 290 00:18:37,792 --> 00:18:43,734 artificial AI engineers. This in some ways has the 291 00:18:43,742 --> 00:18:46,906 opportunity to massively democratize and and reduce the 292 00:18:46,914 --> 00:18:49,850 need for you to be some specialist. 293 00:18:49,858 --> 00:18:54,502 An example I gave is I remember when I was an engineer that 294 00:18:54,510 --> 00:18:56,840 being a CAD expert was a challenge. 295 00:18:56,848 --> 00:19:00,780 You would have to learn how to use NX or CATIA or SolidWorks, 296 00:19:00,788 --> 00:19:04,350 and frankly, if you didn't know how to use those, you couldn't 297 00:19:04,358 --> 00:19:06,000 use it. You would have to go and find 298 00:19:06,008 --> 00:19:07,540 somebody, have to pay someone to do it. 299 00:19:07,548 --> 00:19:10,615 And I personally found it a blocker actually, because I knew 300 00:19:10,615 --> 00:19:14,004 what I wanted to do, I just couldn't do it. 301 00:19:14,012 --> 00:19:19,435 Now Fast forward and by the way, similar extent to some software 302 00:19:19,435 --> 00:19:21,680 tools where you know CFD, right? 303 00:19:21,680 --> 00:19:23,720 I understand the theory, I understand what I want to do. 304 00:19:23,720 --> 00:19:25,960 I just don't know how to use the code or I don't have to script 305 00:19:25,960 --> 00:19:29,040 it or whatever. As long as the agent knows how 306 00:19:29,040 --> 00:19:32,600 to do it. I can say, you know, create this 307 00:19:32,680 --> 00:19:37,200 in this CAD package, set up the simulation, this tool. 308 00:19:38,280 --> 00:19:41,640 It's kind of interesting then that it relies on a good skills 309 00:19:42,280 --> 00:19:45,560 capability. It relies on those tools being 310 00:19:45,560 --> 00:19:48,440 open to agents. And that is a very interesting 311 00:19:48,440 --> 00:19:53,440 point, which is will companies embrace this, you know, agent 312 00:19:53,440 --> 00:19:57,240 first approach or will they try to put more guard rails and and 313 00:19:57,240 --> 00:20:00,880 even block people to use those tools via agents? 314 00:20:00,880 --> 00:20:03,720 Will there be some terms and conditions and licenses? 315 00:20:04,720 --> 00:20:06,840 And and that brings up another point around licenses. 316 00:20:06,840 --> 00:20:10,360 A lot of licenses are based around a human having a seat on 317 00:20:10,360 --> 00:20:15,640 a on a machine, and just like cloud with the ability to scale 318 00:20:15,640 --> 00:20:18,440 disrupted the per core license model. 319 00:20:18,920 --> 00:20:22,375 And really wanted it more to be like a just a power session 320 00:20:22,375 --> 00:20:24,960 like like STAR-CCM+ has with its POD license. 321 00:20:25,680 --> 00:20:29,200 Finally, and I know this is rambling on a little bit, but 322 00:20:29,200 --> 00:20:32,464 I'm just dumping all my thoughts down here, what about open 323 00:20:32,464 --> 00:20:33,867 code? Open source codes? 324 00:20:33,875 --> 00:20:37,936 Much of the challenge was around support and capability and 325 00:20:37,944 --> 00:20:41,808 validation. Is it not the case now with 326 00:20:41,816 --> 00:20:46,080 coding agents and the power of LLMs that these codes actually 327 00:20:46,088 --> 00:20:51,735 could become very capable and and could be used more easily in 328 00:20:51,735 --> 00:20:55,224 a workflow instead of a commercial tool because the 329 00:20:55,232 --> 00:20:59,579 agent can write and script and validate and go off and do test 330 00:20:59,579 --> 00:21:05,483 suites as much as possible? So I'm not saying at all that 331 00:21:05,491 --> 00:21:10,094 you will see a complete wipe out of all the current CAE 332 00:21:10,102 --> 00:21:12,375 companies. In fact, it definitely won't be 333 00:21:12,375 --> 00:21:17,816 like that, but I just put this out that there is a very 334 00:21:17,824 --> 00:21:21,360 interesting opening it because these two new technologies 335 00:21:21,368 --> 00:21:23,103 around sort of foundation models. 336 00:21:23,111 --> 00:21:26,380 This idea that you could have simulations that would run in 337 00:21:26,388 --> 00:21:28,846 seconds or minutes instead of many hours. 338 00:21:28,854 --> 00:21:33,156 The idea that if you have such surrogate models and foundation 339 00:21:33,156 --> 00:21:37,720 models tied with agents, you could finally realize the sort 340 00:21:37,728 --> 00:21:41,800 of reverse engineering or design optimization, the automatic 341 00:21:41,808 --> 00:21:46,580 generation of cases. If you have a pipeline that is 342 00:21:46,588 --> 00:21:50,872 able to do that, then you could simply have agents go and 343 00:21:50,880 --> 00:21:54,560 creating optimum geometries for you, running those simulations 344 00:21:54,568 --> 00:21:57,980 autonomously with you being essentially the the 345 00:21:57,988 --> 00:22:00,160 orchestrator. The potential for productivity 346 00:22:00,168 --> 00:22:03,260 is absolutely massive. It is. 347 00:22:03,268 --> 00:22:07,699 It could finally realize some of the dreams of people in the CAE 348 00:22:07,699 --> 00:22:11,612 and the EDA sector around productivity boosts, around 349 00:22:11,620 --> 00:22:16,125 finding designs around people with just domain knowledge 350 00:22:16,133 --> 00:22:18,490 driving the code rather than code knowledge. 351 00:22:18,498 --> 00:22:21,520 And that is important from an education point of view. 352 00:22:21,528 --> 00:22:25,332 I would say what matters more now is not are you really good 353 00:22:25,340 --> 00:22:26,772 at writing Python? No. 354 00:22:26,780 --> 00:22:28,740 Are you really good at using OpenFOAM? 355 00:22:28,748 --> 00:22:31,440 No. Do you understand physics? 356 00:22:31,448 --> 00:22:35,308 Do you understand numerical? But do you understand how to 357 00:22:35,316 --> 00:22:37,899 pose the prompt? Do that is more important? 358 00:22:37,907 --> 00:22:41,210 Do you almost then it starts to become that you need to go 359 00:22:41,218 --> 00:22:44,152 higher up. If it's so easy to run a CFD 360 00:22:44,160 --> 00:22:47,160 code, why don't why don't I start to run an FEA code as 361 00:22:47,168 --> 00:22:49,299 well? Many companies are you are split 362 00:22:49,299 --> 00:22:52,000 if you are a CFD person or an FEA person. 363 00:22:52,008 --> 00:22:54,140 You're split if you're an acoustics person. 364 00:22:54,148 --> 00:22:59,000 Well, now you could become more than engineer who simply is just 365 00:22:59,000 --> 00:23:01,918 driving the tools. That position in a company is 366 00:23:01,926 --> 00:23:03,574 often the sort of chief engineer. 367 00:23:03,582 --> 00:23:07,580 Well, maybe now it's more important that you understand 368 00:23:07,588 --> 00:23:11,130 trade-offs and and larger engineering concepts and doing 369 00:23:11,138 --> 00:23:16,315 this sort of specialist deep down analysis can be more 370 00:23:16,323 --> 00:23:18,715 automated and clearly simplifying what it. 371 00:23:18,723 --> 00:23:21,114 There are many industries where this will happen for many years, 372 00:23:21,114 --> 00:23:25,055 but I'm I'm sort of trying to Fast forward and play a what if 373 00:23:25,055 --> 00:23:27,720 scenario to, to to all this point. 374 00:23:27,728 --> 00:23:29,743 It raises many interesting questions. 375 00:23:29,751 --> 00:23:34,596 Therefore, on the if the LLMs and the agents are key. 376 00:23:34,604 --> 00:23:37,480 It raises questions about traceability around sovereignty. 377 00:23:37,480 --> 00:23:39,880 Who owns these models? 378 00:23:39,880 --> 00:23:41,920 Where these models done? Should you be developing your 379 00:23:41,920 --> 00:23:44,480 own models? Can you take them off the off 380 00:23:44,480 --> 00:23:48,000 the shelf? So many, so many questions. 381 00:23:48,680 --> 00:23:54,320 It, it rises, but I, I find it a fascinating area that I think 382 00:23:54,320 --> 00:23:59,197 will be a key technology driver, an innovation area for, for, 383 00:23:59,197 --> 00:24:03,637 for years to come and something that I'd like to explore over 384 00:24:03,637 --> 00:24:05,734 the course of this, of this season. 385 00:24:05,742 --> 00:24:09,725 I want to look at speak to some people who can give insights 386 00:24:09,733 --> 00:24:14,214 into this on the the the core ideas around sort of foundation 387 00:24:14,214 --> 00:24:16,600 models. Because the reason that 388 00:24:16,608 --> 00:24:20,762 foundation models of surrogates are so important is that if a 389 00:24:20,770 --> 00:24:27,147 tool still takes a day to run the the the agentic sort of 390 00:24:27,155 --> 00:24:31,200 optimization workflows will ultimately be massively hampered 391 00:24:31,200 --> 00:24:37,263 by the actual tool itself. If the tool can match the speed 392 00:24:37,263 --> 00:24:41,120 of the agent, the LLM, then that is what will really have 393 00:24:41,120 --> 00:24:43,414 the breakthrough. So whilst kind of the advantage 394 00:24:43,414 --> 00:24:47,176 of the agent is that, and so as you don't need to use any of 395 00:24:47,184 --> 00:24:49,512 these surrogate model technologies, they're even more 396 00:24:49,512 --> 00:24:55,395 optimum when you when you do. And, and so this is where the 397 00:24:55,403 --> 00:24:59,599 the, the combination comes from. So, yeah, I don't know what the 398 00:24:59,599 --> 00:25:03,544 future will, will, will, will be, but I know that it is 399 00:25:03,552 --> 00:25:06,414 certainly an exciting one. I think this is a great time to 400 00:25:06,414 --> 00:25:10,055 be in the start up space. It's a great time to be 401 00:25:10,063 --> 00:25:13,110 researching, to be in academia, thinking of your PhD, thinking 402 00:25:13,118 --> 00:25:15,010 about what topics you want to do. 403 00:25:15,018 --> 00:25:18,175 It's a challenging time to be an end engineering customer. 404 00:25:18,183 --> 00:25:21,819 You don't know what to pick. It's a challenging time to be an 405 00:25:21,819 --> 00:25:25,392 established ISV, but it's also a great opportunity because you 406 00:25:25,392 --> 00:25:30,480 can potentially help drive your company to bring out the next, 407 00:25:30,488 --> 00:25:34,825 you know, Industry 4.0 or 5.0 or whatever the the number is 408 00:25:34,825 --> 00:25:38,746 today. So I will, I'll leave it there. 409 00:25:38,746 --> 00:25:41,200 I'm sure I didn't cover many 410 00:25:41,200 --> 00:25:45,200 things and I hope you appreciate that this isn't me in any way 411 00:25:45,200 --> 00:25:50,440 diminishing what will be the classic example of any new 412 00:25:50,440 --> 00:25:54,440 technology taking years and years and years to them. 413 00:25:54,880 --> 00:25:57,960 So I'm not saying next year or even 5 years everyone will be 414 00:25:57,960 --> 00:26:01,360 doing this, just like some people are still running on 415 00:26:01,360 --> 00:26:04,800 Windows 3.1 or Windows 95 in some places around the world. 416 00:26:05,160 --> 00:26:10,600 But I am confident, I am definitely confident that if we 417 00:26:10,600 --> 00:26:16,160 look in three years time that the role of agents and, and and 418 00:26:16,160 --> 00:26:18,920 surrogate models on AI will be transformative. 419 00:26:19,040 --> 00:26:22,000 And so please, if you're listening, if you are one of 420 00:26:22,000 --> 00:26:24,880 these people who are still sceptic of AI, I would 421 00:26:25,360 --> 00:26:28,040 respectfully encourage you to to try it. 422 00:26:28,160 --> 00:26:31,840 And I would hope that you will believe me. 423 00:26:32,120 --> 00:26:35,360 The final call out, and I will say though, is in order to 424 00:26:35,360 --> 00:26:40,320 convince the community, we need transparency, we need openness 425 00:26:40,400 --> 00:26:42,680 as much as possible within a commercial mindset. 426 00:26:43,080 --> 00:26:45,680 People who know me know that I have always tried throughout my 427 00:26:45,680 --> 00:26:49,360 career to, to do things in a transparent and open way. 428 00:26:49,360 --> 00:26:51,480 It's part of the reason for doing this podcast is to sort of 429 00:26:51,480 --> 00:26:54,160 share the knowledge. The reason that I've created 430 00:26:54,160 --> 00:26:59,000 these sort of open source data sets is that desire to do it. 431 00:26:59,000 --> 00:27:02,000 The reason for doing these workshops or AutoCFD and others 432 00:27:02,000 --> 00:27:05,240 is to try and share it. I am fully appreciative of the 433 00:27:05,240 --> 00:27:07,680 need for commercialization. It's how the world works. 434 00:27:08,280 --> 00:27:11,520 But I think the one key thing if this agents and AI and 435 00:27:11,520 --> 00:27:13,480 everything is going to be a transformative. 436 00:27:13,920 --> 00:27:16,780 People will only get on board if they can believe it, if they 437 00:27:16,780 --> 00:27:19,080 can see it, if they can see transparency, they can see 438 00:27:19,080 --> 00:27:22,040 publications. I feel like academia and 439 00:27:22,200 --> 00:27:26,080 scientific rigor has a massive part to play in this 440 00:27:26,080 --> 00:27:28,480 transformation. And if it can be done with 441 00:27:28,480 --> 00:27:33,080 scientific rigor at transparency and good science, then I think 442 00:27:33,080 --> 00:27:36,800 the the world will move on. If it's done behind closed doors 443 00:27:36,800 --> 00:27:40,360 with no publications, with just sort of a, you know, black box 444 00:27:40,360 --> 00:27:43,600 solution that that will not convince. 445 00:27:43,600 --> 00:27:46,840 And in fact, that will in some ways be a barrier to this, what 446 00:27:46,840 --> 00:27:49,800 I feel is transformative technology helping. 447 00:27:50,360 --> 00:27:52,480 So with that, I will leave it there. 448 00:27:52,480 --> 00:27:55,160 I look forward to hearing your comments and I hope you'll join 449 00:27:55,160 --> 00:27:58,200 me on this new season. Please listen to the old 450 00:27:58,200 --> 00:28:00,240 episodes. There's quite a few, quite so 451 00:28:00,240 --> 00:28:03,200 many hours, but I hope you'll join for this season. 452 00:28:03,200 --> 00:28:04,720 So with that, thanks for listening.