0:00 Awesome, so yeah, so excited to see you know, being the last keynote, it’s always like you’re just worried that everybody’s going to be catching their plane, so I’m really glad to see everyone here. I’m not sure if you know, I was given this spot to hope to keep people to stay, or because I forgot to mention 3MF last year. So, so 3 MF. So I’m going to talk today about how we can use AI to accelerate our understanding of the universe, and not with sort of more traditional things that now has done, like looking at our data or working on autonomy, but for hardware, like what the people in this room really care about. And that’s the Text to Spaceship vision: how can we get that mission on the launchpad ready to go 10 times faster using AI? And I’m going to start with where we are now, which is making these lightweight evolved structures that are very performant and that are enabling science today, and I’m going to give a look into the future with some, you know, sort of exclusive, hot off the presses de mos that I think you all really like. So it’s been surprising to me, as like a nuts and bolts engineer, to get some like press attention. So I really like to see the kind of like buzzwordy titles that that they come up with, like, what is what is clickbait applied to your work? It’s something fun to think about.

1:17 So, NASA’s evolv structures reduces weight and waiting above the alliteration, that’s a good thing to go for. They look like parts from alien starships. So yeah, I like to use the word aliens. And this one, language mangage NASA’s generative AI builds spaceships. Well, no, not really, that’s that’s ridiculous, right? It’s just making parts. But what if it could? What if we could use AI and language, starting with language, to build spaceships? And our missions do start with a concise language definition of what it has to do, whether it has to return samples from certain spots on mars, or whether it has to measure certain sizes of exoplanet around certain kinds of stars. That’s what it starts with. So that’s what I’m going to talk about today.

2:06 And to calibrate the audience a little bit, as far as like how like AI pill everybody is who’s used the new, like, 01 model from open AI and really tested it? Okay, good. Who’s used generative AI in the last 48 hours to do something, whether personal or for work? Oh yeah, okay, this this is like one of the most AI pilled audien I’ve ever seen. So actually used it last night. I was like, I’m nervous about my keynote but I really want to get a good night’s sleep, like, what do I do? And I’m happy to report I get eight hours of sleep, it gave me excellent advice. And who has heard or seen one of my talks before? So I know, like, how much to rehash. Okay, not not too many folks.

2:46 All right, so just to give some context, we’ve had three years of funding from NASA research so far, and then we just got our funding for the text to spaceship vision. So to calibrate your expectations, that funding started on October 1st and it’s October 3rd and I’m already giving a presentation about it. But no, we’ve got some great things to talk about. My coy is Matt verick, who’s back here, raises his hand. And with this presentation I want to show how AI is transforming mission design, and particularly, from this audience, to get input on how we can use all the great tools that you all are working on to make this vision a reality and get some feedback. I mean, this is like you, you know, my chance to talk to the people who really care about this kind of thing.

3:33 So the work for me really was inspired by science fiction as a kid reading Asimov and heinlin and the AI, or the the classics of science fiction, you know, I dreamed of this future in space where we have these, you know, massive, you know, colonies that we’re living in, and spaceship zooming all over the place. And I was so inspired by that, that I eventually came to work at Nasa, and I got to work on some awesome stuff. I worked on the Hubble Space Telescope servicing missions, and as part of that I ended up being the program manager for space Power Tools this is the pistol grip tool that the astronauts use to fix the Hubble, and also to build the National Space Station. I worked on an element of the space station this is the express Logistics carrier, which is kind of like a glorified Space Station roof rack, but also holds a bunch of cool science payloads. And most recently, I was the Roman Space Telescope instrument carrier manager. So that’s our next telescope in a way it’s like like a Hubble times 100, you can take a Hubble, 100 hubbles type pictures all at once.

4:35 And through this I learned that space is hard and unforgiving. It’s a very difficult environment and you don’t have any opportunity to fix things that go wrong. So the International Space Station is an amazing feat of human engineering, and it’s the second thing brightest thing in the night sky after Earth’s moon. But it took 25 years and over a hundred billion dollars to build, and it sleeps seven. So, you know, that doesn’t scale to that, you know, vision I had of as a child, with like hundreds or thousands of people living in space, right? So I was a bit disappointed by that. And then I saw this quote, you know, kind of way back when, from the Google CEO, that AI is one of the most important things humanity is working on, more profound than, I don’t know, electricity or fire. Good quote, right? I don’t know. And this really set my mind on fire. Here’s like the person that knows the most about technology of anyone in the world, arguably, and he’s saying pay attention. And I thought, what can I do with this, because I’m a mechanical engineer, I’m not an AI researcher, I’m not a programmer.

5:40 So we started out with something humble, and excuse me for the folks that have seen this before, but we started out with the balloon mission. So there’s a place where we test our technology, and this excite balloon mission is really interesting. It’s a football stadium size balloon that lists this telescope over 99% of its atmosphere, and because it’s over Antarctica it can just stare at stars with exop planets around them as the earth spins, so it can get like a really good long measurement. And it had a simple problem, we need to attach this, you know, the scientist came to me, I need to attach this assembly to the back of the telescope, you know, just need a bracket, easy stuff, right? But it has some complicated requirements, you know, it has some basic stuff, like, you know, it has a bolt pattern here, it needs to match up with a bolt pattern there, but it actually has some pretty severe loads. The way these things work is they actually get cut from the balloon and then Plum it, and then a parachute catches them, and then you reuse them. So it’s cool for the reusable, but that parachute shock is a lot of load, and it also has to be very, very stiff, because there’s a cryocooler nearby shaking, and if it couples in with that it’s going to blur your image and affect your science. And because these balloons are really limited and what they can launch, it has to be lightweight.

6:55 So, you know, here’s how and you know, you guys are really going to be familiar with this here’s how a typical human solves the problem. And this is one of our best, this is this guy Drew Jones, he laid out the James web Space Telescope and is currently laying out the habitable world’s Observatory, which is NASA’s next big mission. And I got him because he thought he was going to retire, to make this CAD model for me and compete with AI he, he was the only one willing to do it. And you know, when he designed this, it was stiff enough, easy to manufacture, but I don’t know, there was a miscommunication, it was way too heavy, didn’t meet the mass target. So he came up with another iteration, and here he did what you would expect, he just pocketed it out, right, that’s that’s what people do, a lot of sketching and extruding, but he’s like wicked fast at that kind of stuff. And in this case, you know, he got the MK down, but it didn’t meet the stiffness requirement, and maybe he didn’t quite understand how big it is, cuz, you, you know, those pockets are actually not very good, they’re hard to reach, wasn’t a very machinable part. Said Drew, make it stiffer. He added some Mass back, but now it doesn’t meet the mass of stiff or machine ability requirements. And, you know, I’m like, okay, Drew, you know, I I was doing the FAA and think of something different. So he came up with this, you know, kind of crazy looking thing that actually does meet the mass and stiffness requirements, but it’s pretty unmanufactured, I think even like if you were printing it it would be really hard to get the supports removed. So we wanted something better, right? Like Drew and I, we probably could have spent a couple weeks and and come up with a solution, but having to like highly skilled engineer spend weeks on a balloon bracket is is like not a good use of our time.

8:28 So evolve structures, you know, it’s basically like topology optimization with a layer of NASA requirements put on top of it. So, you know, you just have the preserves, the areas you have to keep, red, the areas where you can’t have material, and you have loads and Manufacturing constraints, and, you know, those are pretty quick to define, these are just simple, simple geometric shapes, and they are fully determined by, you know, NASA requirements and processes. And then you just evolve the optimal design using generative design AI Tools, in this case it’s Fusion 360, and you’ve got a CNC design here and then an addtive design on the other side, and then you just build them with digital manufacturing. So bracket solved. So we can accelerate going from requirements to parts. So what took two Engineers two days and really hustling, doing four iterations, one AI power engineer can do over 100 iterations in an hour and come up with a design that’s lighter, stiffer, stronger, and more manufacturable. And this affects science, that stiffness is really key to not blurring the image, so these are science enabling designs. I think, you know, if you haven’t seen this, you can go and pass it around, onean, if you like. So this is, you know, this is that bracket at CNC machined.

9:51 So what’s really happened here is it’s a paradigm shift, and and a lot of people talked about this it’s a paradigm shift in the job of the engineer. And and I’m not going to go too much into this, but but, you know how it is, you’ve got a designer, you got an FAA person, you got manufacturing people, they bounce it back and forth, some of them are really introverted and don’t give any feedback, the manufacturing people curse the engineers and say that you’re, you know, they don’t know what they’re doing, they don’t know how things are made, and they just give them a high price and a long lead time. And this process takes months and years, and now, you know, we’ve collapsed that. So the AI generates the cad model, generates the design, does the fin an El analysis, does the manufacturing simulation, and then you just make it. And what’s powerful about this is it focuses the engineer on the higher level task, which is really thinking deeply about what the design has to do. And Matt can ATT test this process has changed the way we think. So when we come a project comes to us with an application, we first really want to deeply understand what it is they’re trying to do, and we realize that most of the engineers, they take their tool and they just immediately start designing, sketching, and things, without sometimes really fully understanding what the problem is. And this acceleration, you know, we’ demonstrated going from having the requirements known, the parts in her hand, in less than 48 hours. So Matt went to a conference like this, crowdsourced the requirements for a part, and in 48 hours had these, you know, I think these are like 10in diameter, mat, 10-in diameter, CNC machine, you know, strong, beefy metal Parts in hand.

11:27 So, you know, how do you make these parts? You know, everybody’s probably thinking additive. So I kind of want of, you know, hopefully for the for the current time, change people’s minds a little bit. For 60 applications that we’ve looked at, and we’ve made dozens, actually, C and C has been the right choice, as far as getting a high tolerance part, getting a strong part, and cost and schedule that we want. Additive manufacturing for Metals is is coming along, but as things are now, we’re mostly using CNC. So, you know, digital manufacturing, of course, is just going from your CAD model or your 3mf to your, you know, your G-Code that the robot then runs. So, you know, you see, you know, 5x CNC, lazy powder bed, you guys know all this stuff. Something you guys might not be aware of is, you know, one of the manufacturers that making a lot of these parts for us is protolabs, and you see these 500 Matt and I walk these Halls, these 500 CNC machines, just all the way to the vanishing point, 500 hos CNC machines, and no machinists and no cam programmers, it’s all automated, you upload your model, you get your parts.

12:34 So this design was so successful that it actually ended up getting sort of, you know, overcome by events, the scientists wanted to do the whole back of the instrument with evolved structures. So you can see, in the green circle, the part kind of got subsumed into this larger part, and I’m happy to report this part has now flown in space and returned to the ground. So here’s the nice video that an intern did that sort of goes over this process, I think it’s only like a minute long. So here you can see, you got the balloon hovering over Antarctica, it’s absolutely gigantic, and we’re zooming in on this Gondola, is what they call it, that hangs below the balloon. You got solar rays telescope, the telescope is like a like a half meter, just to give you a sense of scale, looking back it holds this 50 kg payload, but the part only weighs 2.6 6 kg. So that’s pretty good for taking a 10g shoot shock. And here you can see the red and the green, the red is the areas you have to avoid, the green is the areas you have to keep, obstacles and preserves. And next you’ll see an animation of the design evolving, so takes like 30 some iterations, this is done in Fusion 360, and then, of course, we just we have to do the, you know, validation FAA to make sure it works. Yeah, love I love a good animation, and it gives you a good sense of scale.

14:12 So with this success, we’ve used it on a bunch of missions, and one of the really exciting ones has been Mar sample return, where we’ve got, you know, starting to starting to look like that alien spacecraft that the, you know, the journalists want you to to think about. And in particular I want to highlight this one on the upper right that looks like an alien, like, black tennis racket, and what this does. So Mar sample return, there’s this orbiting sample, the sample is orbiting Mars, and it gets caught by another spacecraft, and it it comes down into this like dark cylinder that you see here, and when it enters the cylinder, this paddle has to quickly close off the opening, but it also has to be strong enough to take the bounce back, because it, you know, bounces back and and not snap off. And none of the human designers were able to come up with the design that had a low enough to inertia to swing over quickly, but was strong enough to take the bounce back. So this technology is already enabling science at Nasa.

15:08 This is another thing that we’re working on, Matt was actually back in the hotel just working on now, this is the Alice evolved Optical bench. So this, you know, somebody called it a tangle of metal, this is CNC machine from a single block, very precise, I don’t think you would want to try to print that thing. And this is where we’re seeing that the okato group folks mention this, where you see this virtual cycle, or virtuous cycle. So we went from a 3.6 kg design to a 600 gam design, and because it’s lighter, you know, our cost estimation algorithms are like a million dollars a kilogram, it’s cheaper, and also what mounts it can now be lighter, so that can be cheaper. And this is a cryogenic instrument, it’s oper operating at 50 Kelvin, so any kind of like heat coming into it is really going to degrade the performance. So because it’s lighter, it’s on these flexors, and the flexors can be thinner, so you have less heat loss, which leads to a better performing instrument, because your detector’s colder, and you also have lower surface area, again, better performance and lower cost, because you need less radiators. It combines 10 parts into one part, so you have lower integration costs, and you also have fewer bolts, and every bolted interface in an optical system is a chance for Distortion, and also every bolted interface is a thermal insulation, because flow heat doesn’t flow well. So having this be one piece means it’s more homogeneous thermally, again leading to better performance.

16:38 And a lot of questions, you know, lot of times we get the question with AI in general, but with this in specifically, how do you know the AI didn’t make a mistake? Well, how do you know that the hundreds or thousands of people that are working on your project didn’t make a mistake? NASA has processes for all this, testing, analysis, ways to ensure that things are going to work when they’re on orb it, and we can reuse those, to the extent that I’ve seen so far. So we can valid it by test, and we can do it faster with this evolved structures approach. And I really like this quote, one good test is wor a thousand expert opinions, and it’s really easy to get expert opinions, and also really expensive, and a lot of times it’s cheaper to just do the test. So, of course, this is like, you know, this technology is being picked up by the commercial industry, Divergent Auto, if anybody’s heard of that, and the aado group, which gave a great talk just earlier today.

17:33 So we’ve been so busy with applications, because of the power of this technology, that we developed an input form. So we talk to the teams that have an application, and we have, like, a very, you know, specific checklist to do, you know, figure out what, you know, each project needs, what are their requirements, and document those very concisely. And then, once we have those requirements, we have this evolved structures guide, which is is now public, public, that tells you how to map those requirements over to your, you know, Fusion 360 entry. And when chat came along, we’re like, wow, this can do amazing things with text, what if we could just take the requirements that were input into this form and make the part, and skip the cad and fem, and just do it directly. And so we were working with a company called Sara, and I think we got some Sara folks here still, and we had two genius interns, in a week, were able to get a minimum viable product for this working, text to structure. Basically, what they did is took that document about how you take the requirements and encode it, and they automated it, they turned it into code, and once you’ve solved something with code, it’s solved forever, right? And then you can turn that code just into a code block, text to structure, and now with that code block, you can make all kinds of stuff.

18:56 So the a15 bracket on the left, you just see sort of text input about where the Fasteners are and how they’re oriented, what the mass is, what kind of loads you have, but you can use that same text block to make this design which holds an x-ray detector, or to make this magnetometer Boom for the lunar Gateway space station, or to make this one we’re going to come back to this one, the inputs from this didn’t come from Human, these were AI inputs, we’ll see that later. This application is a secondary mirror structure for the habitable worlds Observatory, and it’s 16 M long, but you can use the same code because it doesn’t care, it just cares about the inputs.

19:39 And we even had some people, and raise your hand if you get a chance to scan the QR code. Okay, a few of you. So we were handing out the QR code yesterday, and you can actually go on your phone and you can design your structure just on your phone. So we’ve, you know, these are things that people put in yesterday, and then I put into the presentation last night. Brooke Chase did a bracket, and I think Liz I don’t know, Liz, are you still here? No, maybe not. So she wins the I think the the competition for the most interesting looking bracket. And, you know, it gives not only does it give you, you know, a real B that you can go just make with, you know, protolabs or print it gives you a stress report, tells you what the stresses are for the different load cases, gives you the igen frequencies, which is what Germans call modes, and you can even do bold analysis. And then you can also Imagine, like, being able to loop back around and and keep iterating the design all automatically.

20:39 And now that you can imagine text to structure, we can start to think about what else can you do like this. So we know we can do Optical mounts, we can do flexors, we can do Optics themselves, we can do component selection, which you’re going to see in a little bit. And then you can start to to combine these into evolved systems, using just either code to combine them together or using AI agents. So, you know, if you add structures and circuit boards and components, you can get automated electronic boxes, that would really speed things up. You can get more complex assemblies, so things like Trust Systems this is a Trust Systems for habital worlds, it looks small here but it’s actually like 5 meters across. And you can even go after really complicated things like Optical benches, by combining structures, your Optical mounts, your flexures, your Optics, and other components, to build these evolved systems automatically.

21:39 And now that you can sort of Imagine Building Systems automatically, you can really start to think about text to spaceship. How can you combine this to do the whole thing? Well, how do people know what to build, or how does the AI know what to build, or the AI combin human teams the requirements for these systems are all defined in text requirements documents, and at Nasa we call those the level four, level five requirements. But where do those come from? Those come from other text documents, the level 3es, the level twos, and these are roughly, you know, decomposed to align with NASA’s, you know, organization, the different human groups that do things. And at the top you have science, that concise definition of what the mission has to do, the, you know, black hole sizes it has to measure, or exoplanets, or whatever. And going from knowing what your requirements are to your sub system requirements, that’s systems engineering, right? That’s how something like the space station, which is more complex than any one human could possibly hold in their mind, we can get that to work with systems engineering. So requirements decomposition is something that, you know, large language models expect to be very good at. And then, once you have those designs, you go back up the tree with what we call VNV, or verification and validation. So verification is making sure you’ve met the requirements in that requirement document, and validation is making sure those requirements actually give you the mission that you want.

23:10 So it seems easy easy enough, right, but, you know, obviously lots of missing puzzle pieces. So, you know, we’re mostly focused on mechanical, and a little bit of electrical stuff, but we’re missing thermal, we’re missing Power Systems, we’re missing attitude control, we’re missing mechanisms. So there’s a lot more work that needs to be done, but the good news is, as we heard today, there are a lot of people working on this. We’ve got, you know, big companies like Autodesk and Seaman working on it, and lots of small companies that folks haven’t heard of, and, if you know, there are a lot more people just today that I know belong on the list, so so please email me and and I’ll add you. So today I’ve talked about Autodesk Fusion 360, I’ve talked about protolabs, who’s making a lot of our parts, I’ve talked about Sara, and hexagon is the m andate solver, is actually doing the the the top op inside of Sara. But now I want to talk about two small companies that are really interesting, because they’re they’re AI powered at their core, and that’s Celadon and Zoo.

24:14 Has anybody heard of zoo before? I’m just curious. Okay, that’s pretty good, mostly, for some reason, people in that corner. So zoo is a both a CAD program and an API, it’s a very, very new company, and it does Tex toad, so you describe what you want, like, you know, a plate with some holes, and it could be you describing it, or it could be a lar languin model that’s that’s describing it. And not only does it make the part, it writes the code that makes the part, which you can then edit. So here you see, for like a simple plate with holes, you know, very hard to see with this particular projector, but, you know, you just edit the code, and they’re very quickly going up the level of complexity in the parts they can make. Here’s a a more complex part, again, you can very easily go in and change it. And this is also, as I mentioned, you can call this via an API, I know the senera folks actually have some demos of that too. And then, even more complicated, here we’re making an entire, you know, car wheel, all text to CAD.

25:21 Okay, next we’re going to go into some of the seladon demos. We have Chris and Jared from seladon here, they were kind enough to put these together in in Rapid Pace here for us. So the first one, you’re going to see Chris prompting into being a Electronics payload, he just starts a project, and then, you know, one thing I notice about Chris, he’s always very polite to AIS, like he knows what’s up. So he prompts into being this Electronics payload, and then you’re going to see it builds a hierarchy of the parts that you would expect to see in something like that. So here’s all the elements, and then he’s going to focus on, you know, in this hierarchy, the sensors that, you know, the AI, you know, this is basically what’s going in in the background, it’s large language models, it’s rag, it’s prompt engineering, it’s agentic systems. And then he’s even going to go and go out to the digi key API and pick parts that might work in this, and look at the different, you know, metrics for those parts to help select one, and then go in, look at the spec sheet for that part, and then add that spec sheet to the context of the project, so that it becomes part of, you know, what the AI knows about what it is that you’re building. So you can see how this could be very powerful, and they’re they’re working really rapidly to do things like, you know, make sure all the components work together, you know, have different iterations of this, using different sets of Parts.

26:54 So I was really, I was really surprised that this is as far as long as you as you’ve come, and being a mechanical person, I think, oh, we could do this for Optics, we could do it with actuators, right, a lot of, you know, off the-shelf actuators you might want. So this next demo is really, you know, getting a little closer to to to Iron Man Jarvis. So here he’s going to start a project for a small set, and he is going to again very politely prompt into being a small set, with, you know, the sub kinds of subsystems that he wants. And, you know, one of the thing key things we’re going to show with this is, it’s not just an understanding of the hierarchy but a spatial understanding. Okay, so here’s the hierarchy of the different elements of the small sat, and he’s going to go into the frame, and next he’s going to go into a spatial view of it. And this is something I think is going to surprise people, that large language models are capable of SP understanding already, believe they’re using Sonet 3.5. And the other really amazing thing about this is, Chris got this working in four hours in 3js, which is a programming language he does not know. So for the folks that are AI Skeptics, this is the power of AI.

28:16 So now he’s going to prompt into being some bolts on it, and you’re going to see, just using its intuition, it’s going to actually pick pretty good locations for those bolts, you know, it put them in weird places, and it hasn’t done anything weird, like, you know, move any of the components outside of the volume. And then you’re going to see him prompt some more bolts, and if you can see where this is going so now we know the volume of that payload, we know the mass of the payload, it came up with an estimate, we know the location and size of the bolts, you can send that to text to structure, just write it out as a Json, send it to text to structure, evolve the optimal structure for that application, and then bring it back in. And this is just, you know, this is again really hot off the presses, you know, minimum viable product, but I think you can start to get an idea of how you can build complex systems really quick, in a way that’s like, you know, Jarvis, like if you are familiar with Iron Iron Man. Yeah, really, really amazing stuff.

29:25 So now I think the text spaceship vision starts to come more into Focus, how we start with science questions and curiosity, launch a mission, and Achieve new understanding. Now I’m going to dive more into this Mission bucket, because there’s a lot going on in there, and you might be surprised to know that for our missions we really only do four full iterations of the system, four times, and each iteration gets more detailed. The first iteration is called prease a, and it ends in a, you know, know physical document that’s just text and images, that’s a proposal that’s reviewed by experts, if it looks like it’s feasible and something that is wanted, we’ll go to phase a. Phase a, your team gets bigger, you get more Engineers, you add more detail, again, you write it up in a report, and that ends up being the mission concept review. The third iteration is the preliminary design, again, team getting bigger, getting more more analysis. And then the fourth iteration is final design, phase C, where you actually, you know, get all the drawings and go forth and start building it, and then you do a in assembly, integration and test, and launch it. But this takes years or decades, you know, for a complex Mission like web or something, decades, and it takes hundreds or thousands of highly trained people to do, and that really puts, you know, a break on what we’re able to achieve and our understanding.

30:55 Similar to evolved structures, what if we can break that Paradigm, what if we can do more iterations using AI. So obviously there’s a lot going on in this box, but, you know, what if we can do an iteration in a month, what if we can do an iteration in a week, and then it puts out reports and the humans go in, much like you saw Chris doing with seladon, and you tweak it, you change the prompting, you make it more clear about what you want. So everything starts with human curiosity, and that’s the science goals, and right now we’re using llms and rag, retrieval augmented generation, to do literature review, to help us, you know, understand what are the right questions to ask, and everything that I show, as we go around here, we’re actively working on. And then you can decompose the requirements, as I already talked about, using large language models, and you can also create the systems models, like you saw with the Celadon Da Vinci demo, and you can write the documents, and this is something that we’re again doing already, writing systems documents.

32:06 And then you can work on the systems design, you can use large language models to drive the component design and selection, like we showed, you can use them to provide expert knowledge, and you can use reinforcement learning systems to work on, you know, complex multi-objective optimizations for the component design, we talked about this little bit, text to whatever that might be, generative design, it might be, you know, genetic algorithms, it might just be just straight up Automation. And you can also use large language models to drive your C tools, right, I think the age of having C tools that have hundreds of little menus and buttons and weird names for things is going to come to an end, where we just, you know, go much more quickly to your intention to what the tool is doing. So verification and validation, lots of folks working on surrogate models, all different kinds of surrogate models for all different domains, and also, at this point, you know, you would want to write out more documents. And finally, when you go to build and test things, you have some uncertainty quantification, you know, maybe you don’t have enough, you know, data for certain simulations that have been identified, you go off and build things, and you also use AI for manufacturing, which some people talked about here, and Matt and I were recently at imts in Chicago, which is this huge, like, 80,000 person manufacturing show, and there are a lot of people thinking about this. And then closing the loop, you have large language models writing out reports and helping you understand, you know, can you reach those science goals, were you too ambitious, were you’re not ambitious enough.

33:46 So the future work, over the next few years we’re going to work from the bottom up, developing evolved systems, getting that text to structure really robust, and we already have, you know, applications that we can we can use to make that robust. You know, text to trust for habitable worlds is really important to us, habitable worlds is like this height of, like, a three-story house, it’s gigantic, needs lots of trusses, we think we can accelerate that. And then text to, you know, circuit, text to flexure, we need to get all those developed as well. And really importantly, we got to work with fabrication, because right now that’s a very, like, human Centric sort of sticking point, so we’d like to work with with Fabricators to have apis, like, wouldn’t it be great if, when you’re designing that part in, you know, automated design, you can have it go out to an API, see if there’s any manufacturing constraints that you’re breaking, and then feed that directly back into to your design. So also working from the top down, so here we’re going to understand the full stack of what mission development takes, and look at places where you can Infuse AI, both places where there’s, you know, low hang fruit, can pick where AI is an obvious win, and places that, you know, maybe have a big impact that nobody’s really looking at yet. So to conclude, AI is radically changing the way that we develop missions, and I think it’s not only going to accelerate our progress, but it’s also going to let us find solutions that have game-changing performance that enable science. And I think this group particularly, you know, can help us with this vision, find things that might plug into it, and help us accelerate our understanding of the universe.