A Forward Look on AI in A&D Part One – Transcript
In this episode of Talking Aerospace Today, Todd Tuthill is joined by Barclay Brown, Associate Director for AI research for Collins Aerospace and leader for the AI Systems Working Group at INCOSE, to take a more forward-looking vision than usual on the potential for AI in the aerospace and defense industry.
Patty Russo: Greetings and welcome to another episode of Talking Aerospace Today from Siemens Digital Industries software. I’m Patty Russo. I’m responsible for global marketing for aerospace and defense vertical here at Siemens. Thank you for joining our conversation. In our previous episode, we covered the final stage of digital transformation maturity. This references is a five-step framework that can be used to assess where engineering and manufacturing companies across our industry and others are in the process of advancing their digital transformation technology and processes. And it helps to identify where they need to go to achieve business goals. Before we introduce today’s topic and guests, let’s do a brief recap of our coverage of this topic.
Patty Russo: Our conversations about digital transformation maturity began in July of 2023 and we covered the drivers for companies in the A&D industry that require them to go beyond where they currently are in the adoption of digital transformation technologies. We explored the advanced stages of the framework, including automation, generative design and closed loop optimization. We also discussed the role AI will play in that journey. Today we want to look forward to what an aerospace industry transformed by AI might look like in the future. Joining us, as always, is our VP of A&D here at Siemens digital industry software, Todd Tuthill. Welcome, Todd.
Todd Tuthill: Hey, Patty, great to be here. Good to talk to you all today. Can’t wait for this interesting, exciting topic.
Patty Russo: Likewise, and we have a special guest today, Barclay Brown, Associate Director for AI Research for Collins Aerospace. Barclay is also the leader for the AI Systems Working Group at INCOSE. Welcome to you, Barclay.
Barclay Brown: Hey, Patty. Thanks for having me, and hey, good to see you again, Todd.
Todd Tuthill: Yeah. Good to see you. And if I could say, Barclay, I really appreciate you taking the time to join us for our listeners. Barclay is certainly an expert in the industry in aerospace, in MBSE and in AI. So he’s just the perfect person to join us today. So, Barclay, I really appreciate having you on the podcast today.
Patty Russo: Yeah, I’m also looking forward to the conversation and excited to be a part of it. So let’s get started. Todd, you have made a few bold statements in the recent past about the future of AI for aerospace. Wondering if you could please reiterate a couple of those statements and maybe share a little bit about why you see AI being so important for the aerospace industry.
Todd Tuthill: Sure. Last fall, publicly, I said by 2033, ten years from last fall, the way we design aerospace products will completely change. That change will be driven by AI. And I said by then 2023 there’ll be two types of aerospace companies. Ones who’ve embraced AI and ones who are no longer in business. And I still stand by that statement, except I think I may want to adjust it just a bit. I’m not sure it will take that long, and here’s a couple of reasons why. If you’ve looked at the recent news and just think about the market cap of NVIDIA. In October of 2022, Nvidia’s market cap was 300 billion. As of just a few days ago, early March, its market cap was 2 trillion. It’s the fastest rise from 1 billion to two billion, $1 trillion to $2 trillion, I believe of any company in history. And that’s just an indication of the impact that AI is having.
Todd Tuthill: And of course, you know, I went back to October 2022 on purpose, because we all know ChatGPT launched in November of 2022 and that’s what kind of started that big rise in NVIDIA. And I think if a lot of people in, in the tech community and in maybe outside the tech community would say, well, AI was invented in November of 2022 with ChatGPT. Well, no, I don’t think so. AI has been around a long time. We may talk about that, but talk about the future. But as I think about the NVIDIA, as I think about things I saw at CES all the growth in other companies outside of aerospace that, quite frankly, I think many of those companies in consumer goods and entertainment and creative arts are actually ahead of aerospace in many cases in adoption of AI.
Todd Tuthill: And when I think about that what’s happening with NVIDIA, what’s happening in consumer goods and other places, I think that 2033 prediction may have been too long. Maybe I need to adjust that, come back a few years. That’s where it’s at, and I think about that because AI is so important for aerospace. It offers so much promise in terms of helping us with all of our problems with workforce, helping us produce products faster and. of course, even making a better environment to attract new engineers, and I’m sure those are some of the topics we’ll get to today in the podcast.
Patty Russo: Yeah, bold statements for sure. Lots of interesting and dynamic changes in the, in our industry and of course elsewhere. So thank you for that look back. So now let’s take a look forward. Barclay, would you give us a look into your perspective on the future of product development? In particular, in complex low volume industries like aerospace. How will AI change the way engineers work in these industries in the 10 years maybe that Todd mentioned or five years or even two to three years?
Barclay Brown: Yeah. Well, first, I completely agree with Todd that it’s going to be completely different in 10 years, but it’s not going to all happen in that ninth year, right? It’s going to be these little steps all the way along. And I think the steps are going to be so small as to be kind of imperceptible. I mean, every once in a while, we have a big bump. You know, when ChatGPT came out, everybody was inspired. You know, it’s like, oh, look what’s now possible. And it was like this big jump. But now, you know, every day we’re finding new ways to apply it. Every day. We’re trying to figure out how to apply it in sensitive industries and things like this because ChatGPT, as you know, is a public service, right? And not something that we use inside of our company. So what do you do with that?
Barclay Brown: So, but if you look way out, I want to go farther into the future than Todd’s willing to go. I’m going to go several hundred years into the future. Are you ready? Right? And so this, of course, is the world of science fiction. And I’m always intrigued by how science fiction depicts engineering in the sci-fi future. Right? Say Star Trek and hundreds of years in the future. Right? And one thing that’s interesting is there still are engineers, right? There still are people designing new products and new things, right, new ships and everything. And I’ll tell you one little story from this. There’s a Star Trek series called Voyager. You probably remember, and at one point, voyagers out there in space all on its own and they decide just in their starship to design this little new portable kind of a fighter, kind of a small ship, right? And they’re going to design and build it, right, right there.
Barclay Brown: So you watch what does the engineering process look like, that they go through to design this new little ship that they end up calling the Delta Flyer. Well, one day they walk into the holodeck, so you remember the holodeck, right? And one of the engineers says, show me the current design for the Delta Flyer and poof, there’s this holographic image of the whole ship. And they look at it and they go. Mm-hmm. Shorten the wings by half a meter. Foop. There’s the change. Right. OK, now let’s add a vertical tailfin. Whoop. There it is. Right. And so they’re sort of interacting with this design in a very fluid way, merging together all of the different disciplines because as they change the wing size that affects the characteristics. If it were flying in the atmosphere with air right, I guess fins don’t matter if you’re flying in space, but you know they were taking in it’s kind of this holistic approach.
Barclay Brown: And you see the same thing with Tony Stark and Iron Man and the way he’s designing the suit with Jarvis’s assistance and so on. So I think AI is going to be in it through all of this. And what it should do somewhere, hopefully before several hundred years, is make the work of engineers much more powerful. They’ll be able to do much more work in less time, more creatively, and have more fun.
Patty Russo: That’s where I come in and say, “Beam me up, Scotty.” Right? Great analogy, Barclay. One that no doubt resonates with this audience. And what you just said, I think should hit home. Hopefully hits home with those in our audience, because what you’re suggesting is it will allow engineers to do the creative stuff that engineers love to do, and I’m sure some would like to be at that point right now and for others, though, the reality of AI might be just a little hard to conceive. So, what do you think, Barclay are some of the most significant challenges in maybe not getting there a hundred years from now, but getting there in the short term? Or put differently, what do you think might hold industry leaders back from embracing or adopting AI?
Barclay Brown: Well, it’s definitely a challenge, but it’s not a new challenge, right? This challenge comes up whenever there’s a new technology, you know, and the people who invent the new technology are often naive, right? And they’re saying, well, this technology is so much better than the past. Everybody will just grab it and run with it. You know, but if you remember when mechanical with CAD, right, computer aided design came along years ago, right, there were still people who thought it was better and easier, and some even claimed faster to draw something using their pencil and their drafting table. Right. But eventually CAD one out because it had so many advantages. Right. But it was processed over some over some years.
Barclay Brown: And trust is a big part of it because you know, as these engineers, let’s go back to the drafting people, right. They had to learn to trust the CAD system. You know that it would draw it correct if they specified it right, it would draw it right. It wouldn’t introduce errors into it. And you have the same challenge right now. So the answer comes along a couple of dimensions. One is just knowledge. A big part of my job right now is trying to educate people on AI and not just at the surface level. At a deeper level, so they really understand how these things really work and that my thought is that that will bring better intuition about what’s possible and what’s not possible, what we should be trying to get AI to do and what,” No, we shouldn’t get AI to do that, but let’s just have the engineers do that.
Barclay Brown: But here’s what the AI could do you know what’s the right mix and how do they work together?” So you hear this buzzword of human machine teaming, you know. What’s the right teaming model for AIS? And it could be more than one AI, right? And more than one people more than one person, you know, teaming together to do an engineering task in the best way possible, that’s going to be the stuff we’re going to have to really work on. And just like we don’t trust an AI out-of-the-box, guess what? We don’t trust people out-of-the-box either, do we? I mean, do you take a new engineer you’ve just hired, have them do something and you just take it and use it? No, no. Somebody checks it. Somebody verifies it. You know, people look at it. There’s reviews on everybody, right? And we’re going to need to treat AI the same way.
Todd Tuthill: Yeah. And Barclay, that’s a very good point and I’d like to dive a little deeper there because when I talk to aerospace customers, that issue of trust always comes up and it comes up at several levels. And I’m going to ask you to dive into a couple different levels. One of those levels you just talked about is can I trust the results? How can I trust the results? And I’m interested in that specifically around this idea of maybe a nondeterministic idea of training models, and if my model changes, what do I have to do to my AI to trust the results? Especially in maybe a certified aircraft kind of thing? So that’s one issue of trust.
Todd Tuthill: The other issue of trust I hear all the time is well, this is my proprietary data. This is my family jewels that that I that I protect in my company. How do I go mix that in the AI algorithm? Where is it going to go, and how do I know that I can trust that data with a piece of software that I may not fully understand what it does? So if you could address those two things for aerospace, that would be great.
Barclay Brown: Yeah. Let’s try if I can remember the two. So let’s start out with the first one, nondeterminism of systems. I don’t hear that term as much as I used to because people are understanding when you understand how deep neural networks work, which are the heart of AI these days, and are also the heart of things like large language models and so on. You see that all they are is massive matrices of numbers and fairly simple calculations are done with those numbers to produce the results. So I often say in what way is that nondeterministic? If I give it exactly the same input under exactly the same conditions, you’re going to get exactly the same answer. Now the point is, you know, as we work with something, even as you’re talking to ChatGPT, you’re not in the same situation each time, right? Because you’ve given it the prompt and it remembers, and things change and stuff like that.
Barclay Brown: And you can also introduce a certain amount of randomness into the response if you want to, which sometimes makes it more creative. But if you dial out all the randomness, you’ll get largely the same thing. So now that said, you’ve probably noticed from working with ChatGPT or Gemini or Claude 3 was just announced yesterday, right? You know it seems it is the first model to beat GPT 4 on all measures they, they claim. Who knows if it’s been independently verified, but by the numbers they’re publishing, it seems to. So things keep moving right? But so the question of do I trust it to give me consistent answers? There’s, you know, there’s enough variability in the prompting, right? If you prompt it just a little bit differently, does it give you a little bit different answer? Well, which is right? Is it lying the first time and telling the truth the second time.
Barclay Brown: Now it doesn’t work that way. It’s going based on your prompt and following on, right. So there’s a lot of work we have to do in that to make those a little more robust so that even for example, poor prompting doesn’t lead it to give you dangerous or misleading results or something like that. All right. And the second one was proprietary data. Yeah, of course. So, that’s really not an AI question. Believe it or not, that turns out to be more about the infrastructure that the AI is running on. So if I go and type my proprietary secret information into ChatGPT literally, or go to openai.com and use ChatGPT, that’s crazy because there’s no guaranteed security. There’s been no vetting of the security of that service.
Barclay Brown: Now OpenAI says we’re not going to use your data, and I’m not saying they’re lying, but there could be leaks. There may not be adequate protections and so forth. So let’s go one step up. Let’s instead of going to OpenAI, we go to Azure OpenAI, which delivers the same model through Microsoft’s Azure cloud platform. Now a lot of companies use that and trust that with some levels of data. Now can you put tech data on there ITAR data on there. I don’t know. Some companies draw that line right? Can you put classified data on that kind of a network? No. Right. So it depends on where is the system and where are we doing it. Now one of the latest trends that’s just coming up in recent weeks here is, can we run large language models on our own workstations?
Barclay Brown: Because as you know on your own workstation you can have any level of data that you’re authorized to have right on your own workstation. And if I can run the AI right here locally with some accelerator box or some AI chip that NVIDIA is going to give me in the next generation and put into my PC, I can run the whole thing right here. So that may be a trend we’re going to see. And I think we will. I think the PC’s, in fact, I think it was Dell, somebody just came out with a PC that has an AI chip in it, and it says, you know, here it is built right into this laptop and it will do AI right here to some level on your PC.
Todd Tuthill: So let’s pull that hardware thread just a little bit. I talked to a company, I believe the name of the company is Delex at CES and their claim to fame was they’re going to put AI on the edge. Can you talk about where you think computing devices, edge devices are going to go in the future? And is AI going to be everywhere in the factory or is it going to be just in one specific spot?
Barclay Brown: Yeah, that’s easy. Everywhere. It’s definitely going to be everywhere, but there still will be differences. You know, like you, if you say, where’s compute going to happen today? Where does compute happen? Well, it happens everywhere. Some compute happens on your own PC, some compute happens in the cloud depending on various things. So AI is going to be same way. It’s going to be everywhere at some level and some things you do will just happen on your PC. If I need you know what GPT 8 will do in 10 years, right or whatever, then that may need to be running on a server somewhere because it may need that kind of horsepower.
Barclay Brown: One thing we still don’t know is what’s the functionality curve going to look like with those big models? Is it going to level off soon? Is it going to increase dramatically for years on end, you know, with multi trillion parameter models or is it going to get smarter? But with smaller models you’re never going to be able to have the ultimate top model running locally on your PC because somebody out there will have a bigger model running on a bigger computer. But does it do more? We don’t know, right? So we’re still exploring what that curve is going to look like. Sam Altman recently was talking about GPT 5, right? They’re not, they haven’t announced GPT 5 or anything, but they’re kind of hinting about it. And he says it won’t be, it will be better than GPT 4 and that sounds like that’s not saying much, but he says no, no, no, it’ll be better than GPT 4 at everything at every kind of task at understanding the questions at reasoning at math, it’ll be better at everything. And he says that’s going to be a game changer.
Todd Tuthill: Maybe we could go back to the holodeck example a little bit and pull on that thread because okay, let’s say, okay, I’m sold. I’m a startup company in aerospace. I love your example, Barclay. I’m even willing to invest. Let’s talk about for a few minutes the kind of technologies that that need to be developed that aren’t here today to get us to a holodeck in 10 years or 20 years or 25 years. What still needs to be out there?
Barclay Brown: Well, you know, in fact I’m giving a talk on digital engineering here at the conference here in a week or two and one of the ideas there is I think digital engineering is going to have to be grown from the ground up. So the first thing I would say to you, Mr. Startup, right, as you’re going to do this, right, is find the state-of-the-art, and not the state-of-the-art like on the ragged edge of maybe it works, maybe it doesn’t, but one level below that that that’s solid, that works, but is advanced and get that. Don’t try to rebuild that yourself, but then figure out how to do your own engineering to integrate things to tie things together. The holodeck example you’re referring to that that sounds fanciful until a few weeks ago when everybody saw the demos of Sora. Have you seen this? This is OpenAI’s new video generator, and it’s shocking the kind of video that it can generate.
Barclay Brown: So if it can generate video like that, what about generating video of the product we’re building, right, and being able to see it.? And so that’s not a stretch anymore. I mean it’s like here and it’s almost, I don’t even have to use the holodeck anymore, because it’s just like Sora, you know. The digital engineering thing, though, is about getting the best and then being willing to integrate yourself, integrate work, figure out a workflow where the CAD is going to do this, and we’re going to feed certain information from a large language model, it’s going to go into this and figuring out that kind of weaving. That’s a lot I think what the digital engineering task is going to be is weaving these things together in ways that fit what you want to accomplish.
Patty Russo: Thank you Todd and Barclay for your great insights on this topic. This really has been an interesting discussion. We are up against time though, and this is a good place to pause. We’ll pick up the second half of this conversation in our next episode. In the meantime, if you’d like to learn more about AI in engineering, Barclay has authored a book on the subject, Engineering Intelligent Systems: Systems Engineering and Design with Artificial Intelligence, Visual Modeling, and Systems Thinking. In a couple weeks, check back to catch the next episode of our podcast where we conclude this conversation. I’m Patty Russo, and as always, thank you for joining us. We look forward to having you back next time on Talking Aerospace Today.
Siemens Digital Industries Software helps organizations of all sizes digitally transform using software, hardware and services from the Siemens Xcelerator business platform. Siemens’ software and the comprehensive digital twin enable companies to optimize their design, engineering and manufacturing processes to turn today’s ideas into the sustainable products of the future. From chips to entire systems, from product to process, across all industries. Siemens Digital Industries Software – Accelerating transformation.
