Ä¢¹½ÊÓƵ

Really appreciated hearing his thoughts on, you know, the freedom to experiment, the freedom to learn, uh, and, and so much that we could do your, your life's journey has just been an inspiration, uh, to hear that today and we we truly appreciate you coming out and and kicking off our space and AI summit. Um, thinking back to, you know, when I came and had my first introduction to NASA, I can remember I came from the national security segment and I was brought in on a project to, uh, map Artemis and uh I said exactly what does that mean? They said, hey, we want you to take a look at the cyber vulnerabilities for end to end, you know, from takeoff, uh, to launch. So you know me and a team got together, got started. And I, I can remember as we did that, uh, we, we drew a lot of things on the board and we was like we're done and Karen Fields, uh, so kindly looked at me and said, Eric, I, I know in the national security segment you send things up but here at NASA we have to bring them back, right? So it was a learning lesson for me. Uh, that you know, everything that we did, we had to make sure that it worked in reverse, uh, because there were human lives at stake, uh, astronauts, uh, such as Mr. Reisman, and I really appreciate that, uh, that was a learning experience for me, but, but now it looks like we have everyone settled for our panel, uh, so we're ready to get into our first panel which is focused on ethically applying AI to critical space missions. I'll turn it over to our moderator, Heather Pringle. The CEO of Space Foundation and a retired general in the United States Air Force. First, I'd like to say, ma'am, thank you for your service. And the stage is yours. Thank you. uh, I'm Heather. Thanks to Garrett uh for that little uh blurb, but, um, it truly is an honor and a pleasure to be here and really grateful uh that you can all join me here today, so. Artificial intelligence comes with a whole series of opportunities as well as challenges we're really excited about what it can do to accelerate, but uh the challenges have been in the news and in the press about the most dire consequences, and that's really where ethics can be brought in to help govern this process and today that's what we'll explore. And with me on stage I have 3 very distinguished individuals and I'm going to introduce them briefly before we get into the conversation here. So immediately to my left we have Mr. Seth Whitworth who is the associate deputy chief of space operations for for cyber and data. And so he's at the headquarters just across the street and he's charged with equipping Space Force guardians with innovative technologies to digitally transform the space force and gain asymmetric advantage over our adversaries. You oversee the development of policy and strategy shaping digital infrastructure, machine learning artificial intelligence and digital workforce requirements. He also serves as a reservist in the Air Force and prior to his time in the Pentagon, you were also part of a consulting company helping local and state governments in other digital transformation. And then to my uh in the middle here we have Doctor Dave Prakash and he's the director of AI governance at Ä¢¹½ÊÓƵ Allen and you developed the processes, policies and infrastructure to ensure Ä¢¹½ÊÓƵ Allen develops and delivers AI safely and responsibly for itself and others. uh, Dave also leads efforts to help clients develop their own AI governance structures and develop a develop and deploy AI responsibly. In your background you have a physician, a medical doctor background, and you also joined the Air Force as a pilot, a test pilot, and you served for 12 years and thank you as well for your service. You have a BA in chemistry from John Hopkins and MD from Upstate Medical University Syracuse, and a master's in business from Stanford University. So thank you, Dave, for being here. And at the end we have Mr. Ramsey Brown, who's CEO of Mission Control AI, a San Francisco-based artificial intelligence company building the synthetic workforce and warfighter, leads a small, mostly human team in realizing and securing tomorrow's advanced agentic autonomy today. Mission control serves the US, UK, and European commercial and defense markets where they deliver synthetic workers end to end autonomous digital twins for any knowledge worker role, and Ramsey is also a senior research associate at Cambridge University in the UK where you focus on. Cyborgization and posthuman economics and holds an MS in computational neuroscience from the University of Southern California. So welcome to the three of you. Thank you all for being here. So the first question is going out to Ramsey. And it's really about why is the topic of ethics in AI important today? What makes this moment unique in the timeline that we see, uh, ethics in AI is not new. It's been around. We've been discussing it for a while, but why is it unique to discuss today? I think it's an interesting question and it's an interesting point in time we find ourselves in having this conversation. February 2025 compared to even three or four months ago for the reasons that we're sitting only a few blocks away from the Pentagon. Which is going to be one of the few places that is not going to be touched by the reversal of executive order 14,110, 13 110, 13 110? 1414 1 time. So under the Biden administration's, uh, functionally AI ethics AI safety bill, which was describing a trustworthy mandate requirement across the. Uh, departments and agencies that are on the CFO list with the exception of the national security community that has now been reversed only a few months ago we could have sat here on the stage with relative confidence and relative security about not only the importance of AI ethics as a concept in a practice space, but that this was a national level priority and that is not the case anymore and yet. As each day passes and we find that the capabilities of machine intelligence systems. Both get deeper, more competent, broader, capable of doing new things and further along in deployment, so actively rubber hitting the road out in real both commercial and mission contexts, we find that this conversation actually is gaining importance even while it appears to be disappearing from our federal discourse. The national security community is one of the exceptions on this for which they get to make up their own rules about what AI ethics and AI safety should be. Um, 2, I think that this is particularly interesting because as we're coming to find most of the AI ethics conversations that the United States and its allies have had have all been largely whiteboard exercises. We ourselves immediately are not per se. Actively engaged in first party hot conflict for which we would have to assumption test our decisions about where ethics and intelligence in the field of warfare have to intersect we haven't actually had to have that rubber hit the road quite yet, so to speak, for ourselves and yet right now there are people making decisions today. About the code they're going to write to train a neural network that will deploy on a small Arduino chip that they'll put on a balsa wood drone that they will fly over Russia or Ukraine as a semi loitering munition that may autonomously make decisions that end in the destruction of human life. That is a hard conversation to have and it is fundamentally a conversation that a butts up against our assumptions about AI ethics, and we're kind of watching it from the sidelines. And yet we spend so much time thinking about this topic, so I find it to be a really compelling thing to think about now because. It is still such the center of what our focus needs to be, and we're watching all of these things play out in real time, in real life in real ways that are no longer whiteboard conversations. And we need to come to find what do we really stand for and believe in when it comes to AI ethics, particularly for strategic engagement, before we find ourselves in those similar positions where our our tires do finally hit the road and we have to see whether or not our assumptions about AI ethics um meet the light of day of hot conflict and we aren't there yet but we we we we should hope to never find ourselves there but in the event that we do, we're gonna want to know what we actually do believe. So what I hear you saying is that uh on the one hand the commercial market might be helpful in determining where some of those lines in ethics are playing it out uh with the end user but for national security purposes we truly want to determine where the lines are. Before we get into something that is more of a hot conflict that would be great and if we have to find out those things on a whiteboard and in plate panels and in nice workshops, the AI ethics remains everybody's favorite golf clap TED Talk concept until we have to admit that there are real tradeoffs here and people are going to have to become comfortable with the real trade offs of do we make that decision or not. Because it abutts up against maybe an ethical percept we have about what our virtues and values are, and we have to determine what is in line and what is out of line, and we haven't had that moment yet. We hope to never have to have that moment, but the more time we spend working on that reality, the better we're gonna find ourselves prepared to make. Good decisions that still adhere this technology to our virtues should that time come. Thanks Ramsey. So let's let's talk a little about how do you define ethics in AI and who determines what is ethical and what is not and I'll start with you, Dave. Right. So There is no single definition. I mean, to talk about ethics, it's a philosophical debate about right and wrong, good and bad, what's fair, what's not fair, what's just, unjust. And that's very much context driven so you know you can make the case in China it's gonna be uh just based on countries, for example, there's gonna be a very different set of, uh, experiences and a different environment that's gonna drive a different set of uh baseline for what's ethical. So by country it's gonna vary by religion it's gonna vary and even in this country I would argue we're having a lot of debates on what's ethical and not. And, and I would go so far as to say, look, I don't think we could get everyone in this room to align on 100 AI or ethical principles because we're all gonna have our own opinions. So. Starting with that framework there's. There's a lot of gray area and so when we apply to AI how I think about it is, uh, you know what are the things that are less debatable. So is an AI system accurate? Is it, uh, resilient? Is it reliable? um is it uh safeguarded against uh adversarial attack? Uh, is there accountability and traceability? Those are the things that I think about as some of the standards for uh uh ethical AI. But part of your question was, you know, who determines what's ethical AI? And you know, on the one hand, a lot of what Ramsey said, you know, these are all great concepts in the void of peacetime. But I always look at it, you know, every problem is fundamentally a people problem and so ethical AI, I offer this, um. You know, or let me add this other context, you know, I've been a part of a lot of conversations with regulators and politicians, especially AI and healthcare, which is where I first cut my teeth in the AI space, and people said, wait, you're building robot doctors, they're going to kill people. There was this fear of AI and so ethics, this whole conversation about ethical AI, I think is in response to the fear of what AI can do when it goes wrong. But I would argue It's not the AI it's not the code, it's not the technology you need to be concerned about so much as the people building it, the incentives and motivations of the people designing this AI because AI is just math, it's numbers. It doesn't have an opinion on right or wrong, uh, and, and so the real concern is, you know, who's building it and I would make the case that, you know, it would, you know, if you're looking to purchase AI or you're looking to develop AI and you're looking to find a partner, you might wanna ask. What's the history for your partner? Do they have a history of exploiting privacy for profit? Do they have a history of deploying AI that's known to cause harm to certain segments of the population? And if the answer to that is yes, then how confident can you be that the AI that they're gonna develop or deploy is going to be ethical or match your particular principles and values so um I think I. We can have the conversation about ethics, but ultimately it's the motives and incentives of the people designing it that I think has the greatest impact. Thank you Dave and Seth, uh, any thoughts on how to define ethics in AI, who determines it and from a military standpoint being a part of the service, you know, uh, what if adversaries have a different approach to ethical AI? Yeah, absolutely this is for sure something we discussed a lot in the small building down the road um but it kind of goes back to the the piece of how do you define ethics for which component of it, right? So like if I. Discuss just from a military perspective, we're not starting from ground zero. We operate under a law of armed conflict. We operate under national security strategy, national defense strategy that advocates for rule-based orders. So how do we apply those principles to the use of the technology and AI, Gen AI, all of those components are just the combination of people, process, and technology. So we're not starting from ground zero when we're evaluating what our ethical use or behavior or how we'll use these tools are. We have a Good strong foundation from which to evaluate and apply them. The DOD has gone out and kind of defined what some of the responsible uses of AIR and a lot of it was already touched on. We want to make sure that it's traceable and accountable. We need to ensure that it's governable, getting back to the, we want to make sure that we're not releasing AI that our adversaries could then potentially used against us. So is it cyber secure? I is it meeting all its marks and then we really have to get after the people. I think there's this scary thing in the world right now that is AI and maybe it's based on Hollywood or movies or wherever it is, but if you just look at it as people process and technology, how do we build, train and equip a workforce to leverage those tools and capabilities from the lowest level of I need to write a report all the way up to the use of lethal force leveraging AI tools. It's it's about the people using it, the people building it, and the partners along the way that make that happen. Well, and I, I think uh the human aspect of this is an important one and so if we are talking about humans in the loop or humans interacting uh with this, what have you seen, Seth? Yeah, the space domain in particular is one that we are not currently experiencing with many people, although you heard some before that have. We're going to continue to experience it through the data, and that's data perceived through sensors or capabilities through all of these pieces, and this massive amount of data is where we can really start to leverage this AI to make sense of it, but again. I, I asked AI, right, we all go, go to this question of, um, what's the worst case scenario? What's the Skynet, what's the whatever I, I was curious what AI thought the most terrifying thing of a human would be, uh, and so I went to EPT and I said, what are you afraid of when we think about humans? And the response I gave was exactly what I'm talking about. It said, I'm afraid of data. I'm afraid of bad data. I'm afraid of not being able to explain a decision that I made and I'm afraid of not being able to understand that process. So if we boil this all down to. The ingestion and evaluation of data and start from there we can then enable for me guardians to go out and train the models and equip those pieces and on the tail end of that we can train them for things to look out for when the model goes awry because I have yet to see a perfect piece of technology that does not have an error of some sort. And so how do we Change our focus and our TTPs to ensure that we're looking out for those things ahead of time. How are we evaluating them as they're being built so we can know where to evaluate further down the line. Dave, you have a lot of experience with human-computer interaction and thoughts on this, so how has the role of humans with the technology been evolving? Sure, and what we're getting at is nothing new or specific to AI. I mean, you know, we've been talking about human systems integration, human factors for decades across so many other industries, and we really need to just take all those lessons learned in, you know, how we build a car, how we create systems for, you know, people to monitor hydroelectric dams, and so one of the key areas here is automation. I think a lot of what AI is going to do is automation. And one day when it's fully trustable to completely automate some processes, I think we can step back, but I think for the near future there's going to be this period of time where there's going to be human machine teaming and how do we get, how do we design systems so that they're uh most adapted to the unique characteristics of humans and so that's human factors and the example with automation I always think about flying and autopilot systems. I mean you've all seen the news about. Uh, these complex aircraft now that are on autopilot 95% of the time and when something goes wrong, pilots sometimes struggle to actually intervene when intervention is necessary and, and there have been studies in the human factors world for years where the less a person. The frequency with which someone interacts with the system as that frequency goes down, the less likely that person is to intervene when intervention is necessary. So this is like your train engineer that sees a fully automated train suddenly malfunction they go wait a minute, the machine's gonna. Fix itself. Let me just wait until the train goes off the tracks and something catastrophic happens. Uh, I like to argue I flew a Cold War bomber with an autopilot system that tried to kill me at least once a week, so I never trusted the system, so I never got complacent, but I think all of you have experienced this. How many times have you navigated somewhere using your GPS and it's somewhere you've been 10 times and each time you get in the car you're like, wait a minute, I need to follow the GPS that's automation buys that's the complacency that kicks in because you're no longer paying attention the way you used to, uh, I, I think many of you. Remember writing directions on the back of an envelope and then having a map in your glove box and then the one time you finally made it there, you know, it's so ingrained in your brain that you don't need to you you remember how and so I think this is what we're gonna see with AI as automation. is still dependent on humans. Uh, we really need to understand the characteristics of humans. Uh, you know, another example is with, uh, self-driving cars. It's there's a reason why we don't do Level 3 automation because we can't trust people to put their phone down and to intervene and take over when the computer vision model in a car can't tell the difference between sunlight glare and a white tractor trailer and you know someone dies, someone has, and then so. I, I, we're not paying enough attention to the human element in the in the teaming today, and I think when we think about space applications, long endurance space flight when you're isolated. You know, potentially for a couple of years at a time and you're using LLMs that are disconnected from the internet, we really need to think about where it can go awry, you know, what kind of testing is involved so that astronauts can have confidence in the system. That it's not going or Astronauts in isolation for 2 years, that's a different subject than an astronaut in, you know, a ground speed zero in a controlled environment and so we really need to think about the human element and design AI to address those human factors. So we've kind of started with a bit of a near term today discussion but Ramsey, you're thinking well past that so where do you see the future of human and AI interaction and what is the importance. Of ethics and all of that. Sure, and I just want to submit to Dave my Waymo account as counter evidence. I live in San Francisco. I take a level 3.5, probably fully autonomous jaguar like 4 times a week where I I hail it on my phone and it politely greets me with my name when I opened the door because the Bluetooth got close enough. And then like magic this thing just takes off through the suburban war zone that is downtown San Francisco and fails to uh strike any of the uh pedestrians, urban ne'er do wells, potholes, uh muni buses, random ad hoc things in its environment. Um, as I'm sitting there, you know, just fumbling around on, on Twitter on my phone, it is wild how fast our baselines drift around trust with autonomous systems and the things that seem. impossibly complicated for an AI system to solve, it does appear we now have the data to suggest because we've been we've been doing this for as a field, the better part of slightly more than half a century and in modern times, so I'd say everything probably you know 2012 cloud revolution to today so we've got a good decade and a half of. Cloud progress under our belt which allowed for really greater data density and on demand compute which is a lot of what kicked off the current trend and revolution and why we're having a panel like this now as opposed to 5 years from now or 5 years ago, uh, we're just finally at this part of the maturity curve. What we found time and time again is that things that we said are categorically off the table for AI to do, especially on autonomy, are entirely downstream of how willing we are to be patient or spend the money or the cost of electricity, the things that we said. Uh, GPT2 couldn't do or shouldn't do, and I do mean those as distinct, couldn't do in the sense of man, this chatbot is trash it barely produces legible English, um, and then the risk community looking at this chat trash bot uh sorry, trash chatbot and saying um. This thing is actually so good we should not publicly release its weights. That would be a national security concern, and that was 4 or 5 iterations ago of language model capability. We looked at and confidently said it was probably going to take us decades to get beyond an absolutely garbage chatbot to a thing that even remotely generated passable human text, and then it turns out decades was really like months. And then when we got to GPT 3 through to chat through to 4 through to 40 through to 40103, uh, now we're finding that we are probably in a good hard fast takeoff scenario towards really good generalized machine intelligence. Everything I just said two years ago would have sounded asinine. You would have booed me off this stage and sent me back to the loony bin that is California if I'd said that in 2023. And yet we find that with training runs with our willingness to sink the cost into the electricity and the fundamental like GPU watt hours that it goes into training these systems, the quality of data we get against them, things that previously seemed science fiction become science fact, and we have now become logistics, and I think about the old four star quote of experts or amateurs talk strategy experts talk logistics. So what's the interesting question around logistics for the future here? Um, to me it is that the interesting future of how humans and sufficiently advanced and autonomous machines interact, um, downstream and apropos of nothing of safety risk. Of employment risk of some of the long tail that we consider really AI safety things around containment, control, alignment, corrigibility, the holy cow, we invented the thing as smart, if not smarter than us, and it runs at an industrial scale. How many of our critical systems can we still count on ranging from culture and psychological warfare containment through to who runs the critical infrastructure of things like the power grid and the things that fall under DHS that keep the country safe. Um, apropos, none of those things. What does all the good outcomes look like and how do humans and machines interact? The interesting question to me becomes that of taste. There are some things that no matter what Dave Prakash is going to be better at than the most sophisticated AI system because they are downstream of the unique collected experience of your life and the things that are very, very hard to trap in tokens and latent space that you can't get into the training data properly for which your unique lived experience provides you strategic edge on particular taste driven decisions. And the way that we end up operating with these machines in that not very far near future is Dave when so many of the daily things that today seem like, yeah, we could barely trust a machine to do that now become as rudimentary as the machines that route the packets of your traffic for which the people on Zoom are listening to this right now like those low we we treat how things like those low level abstraction layers when more and more of our life goes to that. The last thing that you and I or Seth or Heather are doing are things that involve taste. There are things that involve that human discernment and human judgment that no one but Dave could do or no one but Seth and no one but Heather could do. That to me is the interesting question like what is the near and even farther afield future of this look like is. When more and more just becomes logistics with these things, where do we still have our edge and that that edge lies in those things that no one but Heather could accomplish. That to me is an interesting thing. And I think that's a really great jumping off point to discuss applications and how do you see that. Uh, being applied into the space force, where is AI most effective? Where do you want the humans? Where are you going to want more, uh, agent autonomized uh uh activities going on and how are you doing it today? And uh if you would. Um, there was a memo that once put a pause on artificial intelligence in the space force and then you've rescinded that so that AI could continue to be developed. Can you kind of roll that all into an understanding for us? Well, pause, you didn't hear about a pause, did you? For sure I'd love to touch on it. I think we we need to go back for just a quick second to the risk conversation because it is a really unique one to explore when we start talking about space. I think back to Lieutenant Whitworth who operated a satellite, and it was me to one satellite and the partner next to me sat them to one satellite, and we operated that satellite individually. And that made sense at the time because the DOD had the largest constellation on orbit and we were doing just fine. That very quickly exponentially changed as more commercial providers started launching more equipment and then the DOD itself pivoted to more resilient and proliferated architectures. No longer can I have one guardian flying one satellite. There's gonna just be too many satellites and not enough guardians. And so we have to evaluate the risk of of where can we pull guardians out where is that trust level because unlike the other domains, if there's a conflict on land when the conflict ends, you can clean up the land when there's a conflict of the sea, you can clean up the sea. Space is not one we can actively go up and clean out right now. And so what happens if there is kinetic conflict in space and we can no longer use or levy that domain? We have to prevent that first mover or first strike capability. And so where can we start to leverage some of these automations and what trust level is required where I now can have one guardian reviewing 1020, 1000. I don't know what that number is. It's the conversations that we're having. And there is a trust level that the domain will no longer become available because a mistake in the AI caused a repeated chain. And so we're exploring all of those capabilities. We're looking heavily to industry thankfully we can look to some constellations that are already on orbit who are who are leveraging AI capabilities to do this functionality now. If you go out and talk to guardians, I think they would say they, they want a chat like capability. um, we've been doing AI in the space AI and ML just under kind of different contexts and names as we talk about space domain awareness. That's that's probably our number one priority right now is as we bring in more and more data. For more and more providers, how can we churn through, analyze, and make useful decisions on that space. So we're targeting space domain awareness. We're targeting some of the more autonomous pieces, and we can pull out. And then guardians constantly ask me how can I have more time to go focus on operations. I want to use that unique experience that that Guardian gained doing whatever operation that they were doing that I can't train into a model right now to war game and practice and exercise and test. So can I simplify their OPR EPR writing using a chat like capability? Can I uh simplify their summarization of reports Again, Lieutenant Whitworth used to look through a green book and see what happened the day before, and I would read line by line of like each little contact and see what happened. Can I sum that up in a one sentence that AI analyzes and said yesterday was there were no anomalies. You're free to proceed. Save me 30 minutes and now I can go focus on the mission. Yeah, absolutely, and those are the capabilities that we're looking to bring. We want to enable our program offices to work with industry to bring in amazing cots and. Umar solutions. We want to ensure that we're doing that in a cyber secure and thoroughly reviewed and trusted process, but we also want to enable the functionality to target the more, I don't want to say mundane because those things are very important, but where can we free up that time? I appreciate that and uh you know guardians are not quiet when they want a capability that helps them go faster, so I'm glad you're listening to them. I realize I didn't answer your pause question, which was a good question. Uh, I don't wanna dwell on the past, which is probably why I didn't answer it because yes, it happened, uh, leadership at the time decided to take a strategic pause. It was not a ban, it was a pause. Uh, this was before rag-like capabilities existed. There was, excuse me, a whole lot of unknowns, and we were champing ourselves as the innovation service, and there was fear that um. Data would leak or we didn't know those pieces and so yeah we said take a strategic pause. We have done so much since then, since that first memo went out. We held a generative AI challenge where we said let's get after the training aspect. Let's get after the industry aspect. How can we bring people in? We've targeted more use case scenarios. What do you want to use Gen AI or AI capabilities for and how do we get after those pieces. So while we took a pause, we learned a whole lot along the way and we were able to work with the Department of the Air Force and the Department of Defense to reestablish some of those guidelines and ensure that we were moving forward in a secure and way that didn't hamper innovation. Well, and I think that's why forums like this are so important because you mentioned fear and really there's a lot of information or whiteboarding or experimentation that can help increase the human's understanding and the human's adoption of a technology so that it can be employed in a way. That uh is relevant and uh in some senses what you're saying is the only thing that's stopping us from employing it more rapidly uh is ourselves and our own understanding so AI where is AI failure not an option in space either Ramsey or Dave, uh, do you have thoughts on this? You wanna go first? Sure, um, so. The moniker of my organization Mission Control AI is per the old belief that failure is not an option. When we look at what AI failure is made of either in the short term, medium term, long term, and long tail kinds of contexts. Even outside of space in any domain, these become clear and these go from things like ethics compliance responsibility, governance through to we've got to have a talk about the economy at some point to holy cow, I don't want to be turned into a paper clip. It would be prefer, which is a bad inside joke from the AI field of it's very hard. To encode quiet implicit human virtues and human values into machines that have to always do the right thing even when the right thing is nebulous and we don't know what we're asking for and no one really understands what good and evil are anyway, um, those are all hard things to try to figure out on land, on earth. When I think about those in this context of the domain of space, I think that AI actually has a lot to learn from something very unique that space has to figure out. Um, particularly around now the advanced edge of AI that that we work on around autonomy, um. If for whatever reason. Uh, astronauts are incapable of direct communication, uh, with Earth and with ground control, or mission control, about what should be going on at any given point in time or what the correct course of action would be. There have been developed different protocols and ways of thinking around accountability and decision making in slightly more autonomous and disconnected situations. Her point that I think Dave made earlier about what does this mean in terms of really long haul kinds of situations where we might be going quite far from Earth or quite far from communication or incredibly communication degraded environments. Where someone needs to be able to make a decision that is the good and accountable and just and right and effective decision um either in a strategic context or even in a peaceful context, that's actually a problem in space. Like there's a real thing space space has to contend with as a domain. We're now finding that we have to do that in AI too as we continue to build machines that are increasingly autonomous and we give them greater and greater time horizons of autonomy. It is one thing for one of our synthetic workers to wake up and look at my email inbox and make some triage decisions automatically about what I should be alerted to or not, and that whole operation takes only a few moments and I get a nice ping or text message. Hey, you're actually all good this morning. Keep going, man. um, that operation might last a few seconds. What happens when those operations actually last a few minutes where I say look, I actually want you to prep for me um a little bit of background on the panel talk I'm gonna go give on Wednesday. Can you, can you help me prepare for this? And a synthetic goes out and performs some deep research to brief me on uh what I would need to know and that task might take minutes. What happens when those start expanding out to hours and days and weeks of autonomous operation where there's no human supervising that in the loop. These are problems that are not dissimilar to the types of things space has had to figure out, and I think there's going to be a lot of lateral knowledge transfer and overlap from the human factor studies that happen around space to the human factors studies that need to be happening around AI. And what I can guarantee you is these problems will inevitably converge because space is very cold and very hot. There's not a lot of atmosphere whatsoever. There's an incredibly high amount of radiation at certain points in time, none at others, and it is generally a terribly inhospitable place for for. Fleshy things to be in, which is why we spend so much time building really wonderful apparatuses that help us survive up there and uh even though so on uh other other planets or or other celestial objects. All this is to say If you, if you, you really made me make a long bet about the future of space in AI, it's that space becomes a place that is mostly inhabited by AI because it is so inhospitable to life, whether that is entirely digital systems or these are increasingly embodied systems, we already find that there is at least a planet that is entirely inhabited by robots. Mars is entirely inhabited by robots um that we we do send already autonomous things up there. um, the interesting questions become. As The work that we're doing today in increasing levels of of determining degrees of freedom for behaviors and how you achieve your goals using things that are downstream of language models as that starts trickling into what ends up into space. We will see that some of the problems that space has already figured out about how to make good decisions in degraded communication environments are actually to the benefit of AI when it has to go to space. We will find that, Dave, to your point, you've got a language model up there that maybe has not, that doesn't have the ability to go pull resources or get information has to be relatively clever, and the ways we think about resilience and trust and accountability should learn from what space has had to figure out even without AI in that formula. I'll just make one quick comment on that and that's, you know, the, the question was what's critical in AI and so in my research doing AI for supply chain or even trying to think about what's the minimum mission required equipment in a B-52, well, I've learned is it's whatever you don't have when you need it most, that's critical. So, uh, and then I think that's that concept is can't be more true than pretty much any environment related to space efforts. Absolutely well we have uh time for one more question and I'd love to ask each and every one of you what uh advice or what change would you like to see uh from either government leaders, congressional policy makers, or others uh to further advance artificial intelligence ethics and space applications, Seth. Absolutely I think that again going back to people process and technology and really focused on the people and process would be to enable ways to get after the training and the knowledge piece, but for me specifically getting after unlocking industry and how do we enable industry to go solve these problems. A lot of our adversaries do a lot of their work in a vacuum and they don't have the benefit of competition. In these systems to really spurn the innovation that we're going to need moving forward. So focus on the training and demystifying AI, I think would be a great way to say that, but enabling the technology to be developed and leveraged. Thanks. I would say You know, I think we're all here because we want to accelerate AI development and deployment in this domain. And sometimes we think about hey well there's too many restrictions there's too many regulations and then that's just slowing us down um I would argue that there's a counterpoint to that that if we can set expectations set standards, what's required from developers, and I think you'll actually accelerate AI deployment in the long run and what I mean is, uh, you know, right now the theme is moved quickly with AI and that's great if you're talking about what's the next. YouTube video that needs to show up on your phone, but what we're talking about here is something that's high impact and no room for failure and in those environments it's not about just being first it's or it's not just about being first, but it's about being first and being right the first time and each and every time and so uh you know that that's uh. Well, my advocacy for AI governance, not as a source of friction or tedium, but actually a way to build confidence. So it's it's not about moving fast and just hoping nothing goes wrong. It's not about being paralyzed with fear and trepidation of all the things that can go wrong with AI, so you water down the capabilities. So guardians are now burdened with not only do I have to use AI, but I got to do the manual process and double check it, and I'm not actually making any labor savings, but it's actually the third option where AI is a seat belt, not a speed bump, to accelerate delivery of AI for mission critical applications. Can I say one quick thing because you hit on my my favorite topic. I, I work in policy guidance and oversight, so sorry I don't have any money, but uh what I do have is this continued idea that governance should come in the form of enablement and not in control. So where can we implement government governance and policy that enables people within these guard rails and guidelines to to get after these ideas as soon as governance becomes about controlling and stopping and doing whatever the mission fails, we we need to get after an enablement piece. How would you do that? Very carefully, yeah, no, it is, it's a very nuanced answer and to a degree it depends. I strongly believe we need to bring more people to the table um in all of these conversations. It's not just about what I believe or or what my leaders or senior leaders believe it's about pulling in the guardian who's doing the actual work. It's about pulling in industry who's building the tools. It's about pulling in the organizations that are trying to push this. We all have that little unique talent that we can bring and so we need to bring that to the fight and come unified with one voice and that kind of gets back to our earlier point that why this conversation is important right now. So if the burden is. On industry to figure out where those lines are in ethics and you want to do it well before the rubber hits the road or before the conflict gets hot so we need to make sure we're defining our requirements appropriately and. Um, setting them up for success as well, we have just as much of an onus to define what we're looking for to get after those pieces. Is that hard to do in AI? That is hard to, I think it's hard to do in in a lot of these emerging technologies, um, but it can and it and it needs to be done, um, so, so. OK, well, uh, more of that and we wish you a lot of luck there, Seth. So, uh, Ramsey. Any thoughts, advice for congressional policy makers, the government, uh, industry? And, and for everyone, so I think just for general purpose, um, we're lucky that at this point in time. We have a little bit of of of knowledge institutionally built about progress studies. So like the study of human technological progress, there are people who only focus on this. How do we understand the last 1000 years of what drove human civilization forward? What are the mechanisms? What are the trends? And one of the funny things that comes out of progress studies is that a lot of our. The notion of adoption of technology is totally warped by hyper scalers from comparative general purpose technologies like the advent of the steam engine. The first steam engine commercially available in the United Kingdom in something like the 1st 50 years of its existence, sold something like 20 units. In the first half century of there being a steam engine available for purchase on Earth, they could not sell 2 dozen of them. No one knew what to do with a steam engine, and that would from today's perspective of a highly technological society seem nuts like of course you do this and this and this and this but people were staring at me going, well, I don't know do you know what it should do? And you fast forward to today and you, you hear the proponents of AI technologies say things like chat GPT. Uh, and now I guess DeepSeek has been the most adopted high velocity adopted technology ever to go from 0 to some amount of millions of people who've tried it. This number, the velocity keeps picking up on how fast that happens, and yet you chat with people and their answer is, oh yeah, I played with it once or twice. I touched chat GPT once. I tried it. It was OK. I asked it's some impossibly hard problem my friend goaded me on to try it with and it didn't get it right, so obviously this field is a farce and I can go back to, um, situation normal. Go play. That's the imperative that would be the recommendation I would have is everyone, uh, whether you are a personal citizen, private sector, uh, federal sector, defense space, play, go get your hands on the technologies, go start fiddling with them. Even if you do not think of yourself as a computer's person or a data person or an innovator kind of person, if you do not have that kind of self-efficacy story of I'm the type of person who just goes and fiddles with new tech for the heck of it, you don't need to. There's no programming involved in this one. It will gladly take your Visa or American Express and charge you $20 for the good version that has an IQ of $153. Go start playing and as you play, imagine what it means to have a general purpose mind available to you at industrial scale. And how that would change how you do things. That play that hands on, look, I just fiddled with it and then after 45 minutes I kind of got that aha moment of oh wait, what if I showed it this? I wonder what it would do if I X that starts setting in and that's what's needed and that's what's needed at national scale and that's what's needed within. Our different branches to look at and then find pathways to your point of enabling and saying for guardians just get your hands on the stuff and here's pathways to getting your hands on the stuff because no one's going to tell you from the top what you should and shouldn't be doing with this to be efficacious at it. We'll tell you what you shouldn't do with it to not break the rules, but we won't tell you what to do with it to win. You have to figure that out and the only way you figure it out is play. You've just got to go play. I think that's great advice, and if you can get that in the hands of guardians, there's no stopping what you can accomplish. So if you would please audience, please thank the panelists for their time and expertise today. Thank you very much. That was wonderful. Thank you, Heathere, Ramsey, and Dave for those for for such really interesting perspectives on such an intrinsically important topic. Um, one of the things that I think that is very, you know, intriguing about the discussions, uh, specific to space and AI is you hear the terms and. are starting to emerge, uh, from last year to this year's discussion we're hearing conversations about guardians and who they are and astronauts and even the machines are beginning to have a characteristic of their own in this conversation. So as this language continues to evolve and AI and ethics keeps being discussed, it's just, uh, I'm I'm keeping focused on these. Actors that are emerging, especially in the space industry, it's truly fascinating and the conversation keeps evolving and I just love how we are uh focusing on that and how AI can be an enablement um to those actors to get more uh fascinating things done. So, so great job with that. So at this point we're gonna take a 30 minute break, as I previously mentioned, the bathrooms are down the hall to the right. And uh and our demos are set up right in the entryway. Once again you can learn from our technologist uh about integrated intelligent uh situational domain awareness space dowain awareness, and space traffic management, virtual space ground systems, Gen AI at the edge, space space situ space domain awareness processing, and our brilliant new swarms capability. So I'll see you back here in 30 minutes. 

Um, now it looks like our panel three is, uh, ready to go. Uh, we're gonna, uh, the title of this panel is May the facts be with you, AI at the Edge for Faster Processing of onboard Data. Uh, but this panel will be moderated by Tobias Nagle. Uh, thank you so much, uh, for our moderator. We also wanna give a special, uh, round of applause again, uh, Mr. Mr. Nagle for the Air and Space Forces Association. Um, being such a gracious host, uh, you're not only the moderator of this, uh, you're the vice president of strategic communications here at the Air, uh, Space Forces Association, so we just wanna thank you for your hospitality today, agreeing to come and, uh, moderate our session, and now I turn it over to you. Thank you very much. It's a very nice introduction. Uh, it's kind of fun to, to be at someone else's party at our house. Um, so we're talking about AI at the edge and um. Kind of what that means, you know, autonomy or you know some kind of autonomy has been, uh, in outer space since man has been in outer space. I mean, beginning with Sputnik, everything since then, uh, has had some level of autonomy, um, but I, I, when we talk about edge we can talk about lots of different things and I thought I'd start, uh, plan with you and I, I guess I've, I've already blown my opening because I'm supposed to introduce you guys, so let me, let me do that first. Uh, Major General Clint Crozier is director of aerospace, and satellite solutions at Amazon Web Services. Uh, he is also, uh, has been called the architect of the Space Force, um, and he, he really was one of the original people who helped stand up, uh, the, the space force, which is now 5 years old, so he left there and he came to AWS to stand up their space force, uh, so to speak. Um, Omar Hattale is, uh, chief artificial intelligence officer at NASA Goddard. Um, he created the AI strategy for NASA, not just NASA Goddard, I believe. Am I correct? Yeah, well, so we, we led NASA biggerwrit large, uh, so you got the right guy, uh, from NASA here and Dan Wald, uh, is principal of artificial for artificial intelligence at Ä¢¹½ÊÓƵ Allen. Um, and he led the team that developed and deployed the first generative AI large language model into space, so and that's on the ISS. That's right. So, um, again I think it's a great panel to, to, to discuss these things. Clint, let's start with you and just define what we're talking about edge when when. To level set the conversation. Sure, and thanks again to Ä¢¹½ÊÓƵ Allen for inviting us and holding the event today and thanks to AFA for letting us party in your house, as you said. We appreciate that. So I think that's a great way to start when we talk about edge. Edge means a lot of different things to a lot of different people. I'll tell you how we think about it at AWS. At AWS we have more than 100 data centers around the globe and in dozens of different countries. And our data centers are everything exactly that you would think. These are the very large multi-acre campuses with uh you know, acres and acres of storage and compute and large power generation and everything else. The way we think about uh edge is if you're operating inside one of those data centers or plugged directly into that data center, you have the access to everything you would ever want or need. Right, you've got an unlimited capacity as we have all of our data centers are connected. You're living, living in the cloud and and and and every capability you'd ever want. If you're not operating physically inside the data center or directly plugged into the data center, we consider you at the edge. Now there are a lot of different parts of being at the edge. You can be at a narrow edge or, you know, a longer edge, and there are a lot of different. Ways to think about edge, but if you're not inside the data center, we consider it to be an edge capability. And what we learned a long time ago, a decade ago, is it's much more efficient, cost effective, and innovative for us to bring the cloud to our customers rather than their customers have to bring the data to the cloud. And so we're in the business of pushing the cloud, pushing the edge wherever we can. Uh, and when it comes to this organization and this group of people and the customers we all represent, we're talking about edge computing pushing all the way to the edge, and that's operating in space. OK, so Omar. You've got, you're looking at various extensive edge as you uh make your way from uh from the space station to to Mars. Talk a little bit about how that works. So the on the edge is gonna be very essential for us, especially the further we get from Earth, uh, the lag and signal, it makes it imperative for us to be able to to make work with on edge. So right now we're working on several projects, uh, there's a product called BrainStack. And what we're doing is, uh, we're looking at multiple uh uh compute systems, GPUs, TPUs, neuromorphic computers, and getting off the shelf, uh, and making sure they work properly on their low Earth orbit. Because if you go through conventional methods, it might take a long time, a lot of effort, and very costly to be able to do something that's radiation harden, for example, and then by the time you finish this, there might be something more modern, so it might be already obsolete. So we need to experiment getting off the shelf components and see how they work, what is their performance degradation with cosmic radiation and so on. But when we go eventually to Mars, it's very important also we send autonomous systems ahead of humans ahead of astronauts. We need to start having systems that can actually build habitats with 3D printed materials from in situ resources. Uh, it's important for us also to have systems that can go on and get fuels, uh, create hydrogen and oxygen, create the fuel for astronauts to come back. All the stuff has to be done autonomously a very, very high degree of autonomy. But even so you're gonna need that's you're you're not talking about a little bit of compute power you're talking about a ton of compute power. It it could be it could be specialized in nerd. You don't have to have a large language model that is specialized in multiple things. You could have smaller ones specialize in specific concrete tasks that make it much more effective and it doesn't need the bigger infrastructure. But something as simple as as medicine sometimes somebody gets sick, for example, and, and depending on the orbit could take up to 40 minutes between. You send a signal from Mars until you get it back from Earth, so we're creating something called Doctor in the box as well. Uh, imagine if you have medical issues you can have interactions with these, and these, uh, systems are trained specifically on medical domains, and we don't need, for example, the system to be expert in structural design or thermal, uh, components. So that's why I'm saying it has to be components that specified fine tuned or specific domains that will be much more effective for you. So Omar, you were talking about. Uh, uh, basically commercial products. Dan, you've done some work in in that area. How easy or difficult is it to take a modern commercial GPU with it's a big chip, very fine, uh, laser lithography, geometry. How do you make that safe to work in space and reliable in space, or, or is, is, you know. When, when I was writing about this stuff back in the 80s to date myself, that was not even, you know, viewed as a possibility, right? It had to be special materials custom made until very recently, that was right. Um, and I think that our biggest challenge is, uh, the same challenge that Quint was talking about earlier, um, where the big data centers are wonderful if we can afford them, right, if we can afford the size, weight, power, and cosmic radiation hardening. So bigger is always better until it doesn't fit when we're launching things into space, every kilogram every kilogram counts. Every watt counts. So my biggest focus is really about how do we walk down the wattage, and you know we're lucky that right now our test beds are in low Earth orbit, which is largely shielded from the harshest of the radiation that's coming in from the sun. Um, but we are indeed working through some of those challenges today through partnerships with, uh, other startups, uh, uh, and cosmic shielding companies that are, are really working to, uh, take the advances of commercial off the shelf GPUs, which is the the engine behind AI acceleration to parallelize all of that processing and really use it. Right on the edge and so instead of you know a a gigawatt or or a kilowatt or 100 watts or 10 watts we're not talking 10 to 1 watt is what we're really looking for for specialized use cases. I think that's really powerful because when you think about commercial off the shelf or cots, right, it's always a go to strategy to employ cots as much as you can. But think about this, and I love the example that was given in the first panel today where the gentleman talked about the steam engine and only selling 20 copies of the steam engine in the 1st 50 years or so. But think about this. Even what's considered cots today was purpose built 10 years ago or 15 years ago, right? So there are things that are going to appear purpose built today that we are going to turn into cot. Products and by that I mean even when you think about generative AI and AI at AWS, so at AWS we recognize the enormous power requirements and efficiency requirements for computer chips. So we went out and optimized, built, design our own computer chips, training them and inferentiate that are optimized for costs and performance for AI, ML, and Gen AI capabilities. So that was a purpose-built capability optimized for a specific mission that is now considered a COS. We're going to. Through the same evolution in space, you know, it used to be the joke was in any acquisition program, do you want it fast, cheap, or effective, you can pick any two of the three, right? Any acquisition person knew that joke, right now we have to balance sort of the trinity of computational power, energy efficiency, and RAD hardening, right? That's sort of the three things we've got to balance against the standard swap. And so we've got to figure out as much cos as we can use great, but, but the space environment. It being so very different, limited in our bandwidth capability, our connectivity capability, we're going to have to make sure and we're doing this right now today at AWS, by the way, we're going to have to develop some purpose-built things optimized for allowing us to do advanced AIML and generative AI on orbit that may not exist today but are going to be cos for tomorrow and that's part of what we're driving through at AWS with partners like Ä¢¹½ÊÓƵ right now today. So you're going to have data centers that don't need to be connected to. So we have edge capability right now today that operate both in a connected fashion and in a disconnected fashion, right? So we built the software such that if I if I can make a connection back to a data center ET phone home that I can do all my orchestration and data sharing and synchronization, everything else, but if I go into a denied environment, I've got enough compute storage and analytical capability on the edge device that I can continue to operate successfully over a period of time, even in a denied environment. And that's one of the things, by the way, Ä¢¹½ÊÓƵ Allen has developed an interesting capability called Edge Extend, which is a hardware device that operates on AWS software that has been purpose built specifically for that disconnected and operational environment. So when Omar, when you go to Mars. You're going to We're talking long distance. You're going to want to phone home somewhere in between. Are you going to have AWS data centers, um, partway there I mean are will or will all your compute power have to be resident on site? Yeah, so because of the, the signal issues we have to have systems that are self-sufficient on edge, uh, that actually train for certain things, uh, but also it tells you that all these domains, all these fields are moving at such a fast exponential field that wherever we see today is a is a is top state of the art and and. could be obsolete almost so we need to design things with an element of foresight. Where where is technology heading? Not not what we have today, but where do we see from from certain points in in research papers, patents, investments we can kind of have an idea where is technology actually headed and then start designing things based on on what they're gonna be look like and and few in the different time domain. So I think that's gonna be essential, not basing everything on what we have today, even when you're solving problems. Challenges as well we need to adopt a different mentality that have elements of foresight and futuristic elements as well to be able to resolve the issues there. And can I piggyback on that too, by the way, because you asked about data center in space and so again we have to be a little bit careful when we say data center we go to what I talked about earlier, that you know, multi-acre campus somewhere that's not what we mean when we talk about data centers in space, right? The size, weight, and power, you simply can't do that. So how do you build a capability that's optimized for the size, weight, and power available on orbit or on the moon or in Mars, and how do you build it such that it has enough storage capability and enough analytical capability and enough compute power to handle that specific mission? I wouldn't call it a data center, but it certainly is a cloud edge computing capability that allows you to push the edge wherever you need it in orbit, and then you can work your way backwards to the big cloud if you need to. You've got an edge device on Mars. You've got edge computing capability on the moon, then you've got an orbital capacity in space and then reaching back to the Earth. There are people working these kind of capabilities right now today, so you're. Those successive layers become growth capacity, right? I think that one of the biggest problems you have in space is you deploy something that's You know, uh, years away you're not going to be able to send up spare parts, right? So you obviously you have to have redundancy redundancy, but you also have to have growth. You're going to have things that you didn't anticipate capabilities that you're going to want to add 100% of device and, and you know, if you look at space, uh, hardware usually takes a long time to be able to get certified properly and especially when you talk about a mission to Mars, you know, you don't, you cannot send it like in a small amount of time. So we need to look into things that could self-sustain and self develop self upgrade. Uh, in terms of what they're working on, and you know I was looking at even how everything is evolving here in terms of IQ, the, the larger language models. So they estimated cloud, for example, um, at the beginning of cloud 3.0 that it had an IQ level of about 100, and then they estimated in May we're gonna have um uh open AI 4.5, it's gonna have about 155. And the community is actually saying that it's gonna be doubling every 5.9 to 6 months. So just to give you a context, yeah, no, it's, it's incredible. So just the difference, uh, uh, Einstein, for example, had an IQ of 162, 163. It's estimated that that level. And an average person could be 110, you know, approximately 40 points. Look how much difference it made. You know, he was able to come up with a theory of relativity and incredible advances in the quantum field. So imagine now you have a system that's 1000 or 2000 or 3000 IQ levels. So how do we keep evolving the the the resources and capabilities we have according to that the advances we have on Earth? We need to find that bridge to be able to. making progress and advancing at that same speed that we have here on Earth. Well, even even Einstein had a collection of experts that he, you know, compared against and worked with. And so, you know, one of the one of the bright spots that I see in the future is really a collection of experts, right? You talked about a doctor in a box. We've got a technician in a box out there in the demo floor that you can go and take a look at afterwards and and how many more are really necessary and really the answer is. As many as we can send up there and, and with each one of those kind of being smaller and smaller, how many of these collection of boxes can we send up? How many of those need to be at 1000 IQ versus 160 versus 110, right? And and you know, bigger is always better until it doesn't fit. That's really that's really something to come on back to and the collection of experts can sometimes be greater than the sum of the parts. I think digital experts will be something that will be leveraging on substantially. And even something as simple as you're doing, um, a research paper, for example, before you send it for review for peer reviewed, you can have digital expert trained on tens of thousands of research papers on that domain on papers and everything else, and gives you critiques the paper in a way that could have one of the best experts in the world before you even submit it. So I think this is the domain of digital experts are gonna have it transversally across multiple organizations. Can I say something a little embarrassing. My new favorite prompt is two words proofread. OK, so. Lots of technology, lots of opportunities. It's, you know, and, and it's really fun to spend all this time talking about all the cool things that are possible. But when you actually have to acquire something, our acquisition systems are not built for acquiring things that change, right? Everything that has ever been made slower and taken longer to get there has been because of requirements created because it wasn't in the original requirements because we had to go back and change the language that's not what what what what was agreed on. How do we get over that and still move at optimum speed? I'll take this first, right? And I imagine, you know, Omar, you've got a lot to say on this as well. Um, I sometimes introduce myself as a recovering systems engineer. I spent the 1st 10 years of my career, uh, you know, diving deep into that. And so requirements creep is something I fought against diligently for years and sometimes lost. Against um and what I see today with especially with generative AI but but you know with with all sorts as well, including computer vision which we've got another model in the back uh for for us to take a look at as well is how do we trust it we talked we touched on this in the in the ethics panel as well. How do we trust it, um. We have a we have a nondeterministic system that's right most of the time, right, more often than people, and yet we can't necessarily hit all those requirements that we're setting out in in the same way that we would design a software system. So one of the things that. I, I, I think that we need to be focusing on is what are the guard rails to ensure no harm occurs because if you're right, you know, 80% of the time, does that mean you're wrong 20% of the time, or does that mean that you're doing nothing 20% of the time and what is OK and and I think there's a lot of. You know, open space for us to rethink those systems and we really do need to rethink them because designing a nondeterministic system is fundamentally different than designing a deterministic system. You're gonna have to help an ignoramus. Define deterministic and nondeterministic I'll touch on that in the past the talking stick. Determining the deterministic system is basically a set of rules or algorithms. These have been around. It's it's in statements. It's a whole bunch of, you know, collection of rules and statements that says if this passes go down that path, go down this path, go down that path. A nondeterministic system is using something called inference to basically. You know, reason and and come up with the right answer based off of all of the context, all of the all of the training information they have, the, the information they have from the prompts, the information they have about you and this the pre-prompts and the, the, the embeddings that we might have for RAG and and the agent of course capability AI fundamentally different. We, we get questions all the time. Can you guarantee it I'll get the same response next time? No. OK, so how do you, how do you change the, uh, change the acquisition process and, oh, well, changing culture we all know that that's really easy. So, so yeah, so, so you're acquiring systems today for NASA. Let, let me speak and, and I feel qualified to say what I'm about to say as I, uh, tell you that I was the deputy. The programmer and budget chief at Air Force Space Command. I was the chief programmer and budget chief at Air Force Global Strike Command, and I was chief of requirements for the US Air Force. The JSID system, the requirement system, the PPBS, the planning programming, budgeting system within the Department of Defense is the most perfect system you could ever design. If you have an unlimited amount of money and unlimited amount of time to execute it right that and that worked in the 70s and 80s, we didn't have a near peer competitor. There was no pacing threat and we seemed to have more money than we do today. But, but so when you have a near peer competitor, when you have limited resources, limited funding, and now you've got a time element about, you know, staying ahead of the adversary, that system just doesn't work anymore and we have to develop a new acquisition. requirements and development process that takes advantage of things like dev ops and spiral development and we have to procure software as a service rather than hardware as a platform or a weapons system. Those are key changes we have to make. But can we procure software as a service for something that we put on Mars that's supposed to support life? Right, that's we're not so right somebody's going to say, where's the guarantee. Yes, if you procure it as a service rather than as a as a deterministic output, and I'll turn it over to you, but it's got to be you can't define all the requirements up front. You have to define the service or the capability you're looking for up front and then allow the software systems to spiral to keep up with the changing requirements of the mission. Yeah, no, we need to differentiate between the conventional way of certifying and and getting things and and and what's happening today in the incredibly fast evolution. So I think we need to have a look and see where it's, it's very complicated, very, you know, I mean obviously for a reason because when you're dealing with human lives it has to be rigorous and and there's no room for mistakes, but we need to look at other ways, you know, how do we expedite the process maybe that can make a big difference. And just to jump on what Dan was saying about computer vision. Also this is essential to have computer vision on the edge as well. When we send uh systems to be able to to to build habitats with three dimensional 3D printed, uh, whatever happened at the base could affect the entire integrity of these habitats and these structures by having computer vision can make. Tell you and determine if actually there is a mistake or an error at the base and self correct automatically without having to, you know, to continue the building so this could could save you so so much resources and and and save, you know, amazing things. So we need to look into the computer vision as well on the edge and and then we go into something much more essential that um that we need to evaluate in the future also we're talking about robots and everything else but humanoid robots might will be something. Yeah that will be able to send on Mars and uh these humanoid robots will be working hand in hand with humans so we need to look and also into the ethical implications of everything we we look typically at everything that uh ethical fair, transparent, accountable, safe, secure, interpretable, and so on, but let me give you the scenario so you can appreciate how complex these systems might be from an ethical perspective, not only from a technological perspective. So you have a humanoid robot that has on edge systems that enable it to to do incredibly complex issues. This robot, the human robot, is gone with a couple of astronauts on a mission, working on something, and they get into an accident and both of them get injured. And uh and they both have equal uh chances of survival. Which ones do you choose, you know, which one does the human robot select and chooses to come back to the to the base, um, and, and also, you know, we talk about Asimov laws which we abided by forever, you know, that the human might not injure a human and so on, but what if we have on the edge systems humanoid robots will be surgeons on the surface of a distant planet from here. The fact that he's doing an incision on a person that's harming a person and that goes completely against the laws of Asim. So even the most fundamental basic laws that we abided by for long term, they need to be reevaluated, reassessed for the next evolution of these technological advances, and that robot is going to be operating on the edge, so it is an edge device by itself, not just that. And it has to be autonomous or semi-autonomous. I'll I'll piggyback what you said, but I'll take a step back to as was described in the previous panel, AIML, generative AI have become table stakes for our future on orbit systems and capabilities. We have reached the limit of human capacity. To digest petabytes and petabytes of data in real time and make any sort of intelligent decisions about them, right? We've culminated, so we must further embrace AIML and generative AI capabilities for the future. Table stakes non-negotiable right now today, NASA has already started. Porting many of their technical manuals into AWS generative AI capabilities such that you can do a rag chatbot right now in certain parts of NASA and say give me all the specs on a human lander capability and modify X or Y by mass or payload and the system will come back and provide you all that and recommendations right now today at NASA underway. Think about the famous. Scene in excuse me, in the Apollo 13, where you know Houston, we have a problem and you remember they walk into the conference room and they dump out in a cardboard box. Here are all the pieces of hardware that we have available on the on the capsule right now today, and we have to build a recovery system of only what's here right now. Think about porting that into the future in Houston I have a problem becomes Houston, I have a solution because you've got. Autonomous capability on the surface of Mars and you program into your Gen AI capability. Here's all the in-situ resources I have. Here's the storage and compute capability I have. Now go generate me 3 courses of action to solve the problem that I have, and Gen AI Systems will bring back courses of action that will solve whatever challenges you're facing without any human involvement at all. Humans will make the ultimate decision. Necessary fine, but that's the power of where we're at and where the future will be and what advanced cloud computing capability edge device all the way to the edge of Mars and is going to enable us to do. So what does that do to mission control, right? Everybody who's ever watched a movie or anything or watched the lift, you know, a launch, we hear about we see or hear mission control countdown room full of all those people. Will we not need them because this is all gonna be uh native control up there and just you know a 3 person help desk back home so it's it's gonna be an evolutionary process if you look at how many people that were responsible for each console in mission control in the Apollo eras and you see how much, how many people do we have for the space shuttle, how many people do we have in the next vehicle. The more technology advances, I think we're gonna be able to make autonomous decisions or self-sufficient decisions without having humans in the loop. So, uh, the amount of technology advancements is gonna require less people to be involved in the whole cycle, but the amount of people not that's the term will be determined in the future. But if you allow me also to jump on on on on Clint was saying. So, uh, large language models as well, it, it we have to determine where they're good and when they're not good. So they're excellent at, for example, pattern recognition at, uh, mathematical solutions, but they're really, really bad at, uh, social intelligence at abstract reasoning at complex reasoning. So and they keep evolving, so we need to see and understand when can we use these systems for our benefits and when. When we understand that they have a lag or they have a challenge in achieving these results and and for us to understand, you know, where to use them best and where we cannot depend on them. Yeah, so I didn't really like your robot doing surgery very much that didn't excite me. I just really quickly, it's got to be a hybrid because think about it, what did you tell me? It takes 40 minutes round trip. For a signal to move from Mars to the Earth and then back again, that's just a pain. That's just a pain. So the answer is going to be where are those situations where human life is at risk or catastrophic failure of mission is at risk, and if it's less than 40 minutes or an hour round trip, that's likely where you're going to have to default to an autonomous capability to make the decision for you and that's, by the way, could be true of satellites on orbit here in low Earth orbit or GEO. Around the earth, think about a catastrophic collision that's about to take place, and can you use AI? Can you? The answer is yes, but can you use AI to do automated collision avoidance maneuvers and can you do it in a time frame shorter than transmitting all the data to the Earth and having somebody in a command center, you know, manufacture a satellite maneuver, and the answer is yes. With the capabilities we have. Can do that on orbit with AI gen AI software faster and potentially more effective ultimately than what you can do commanding controlling from the ground. I'd love to jump in on that, right? And so like I live in Cape Canaveral and I can tell you I can feel in my bones the exponential acceleration of launches, and I don't mean that I feel that. Literally my chest shakes at 2 in the morning sometimes and it's and it's SpaceX or or or whoever launching and what this means is you know what's happening to our our our mission control our help centers it's really funny um you know we're we're having a scale we're having a scale in at speeds that we've never seen before and it's, it's never gonna move this slow again. And so these are the things that really are needing new ways of thinking in order to accelerate, to scale, to parallelize ourselves, and to augment our operational capacity, uh, our, our ability to make critical decisions with all the information that we have space domain awareness, you know, you just talked about. 5000 objects today. What is it going to be next year? 30,000. What's it going to be the year after that? 60,000, 50,000. It's growing exponentially and it's already too hard. And so what do we do when we move forward? How do we maintain custody of tracks and understand what's going on, you know, throughout the system and then act on it? Some of the. Great conversation of the previous. So one of the discussions that that was that came up earlier was uh the point about manned on man teaming, right? And uh my, my colleagues at the Mitchell Institute talk a lot about man on man teaming because some of, uh, some of this capability isn't necessarily ready to be fully automated, um, and we're not necessarily emotionally, uh, uh, or ethically ready to, to unleash, uh, totally automated weapons, but. Ethical issues deep in space, uh, it's not always going to be an automated solution. It could be an automated set of options, right? I mean, we're to me, ethics is, is really all about can we trust it. Right, and you know, bringing us back to the to the edge compute, you know, uh, topic, can we, can we trust what we're seeing and what, what the satellites are, are looking at and sending down basically 4%, 5% of the data instead of 100% of the data saving us 95% or 96% of our time, um, you know, when we're talking about the scale of things we may not, we not, may not, we may not be saving any time we may simply be actually achieving the mission. And so, so, you know, ethics to me is trust and can we achieve the mission that we have we set out to, to, to do? Can we do what we say is kind of what ethics really means to me and this could actually go into into multiple dimensions of of because you know we're talking about, OK, uh, it's making the best decision, but when these systems start getting maybe orders of magnitude more intellectually capable than humans. Uh, they will probably make a decision that we think it's not the right decision, but in fact it ends up being the best decision possible. And the way I think about it sometimes is like imagine having, you know, quantum mechanics book you're trying to explain it to a 5 year old, you know, that's how, how gap we're gonna have in the future in terms of incredible intellectual capability. These systems we have compared to humans we have our limitations so we'll be, will there be a disconnect between what these systems are telling us and what makes sense and what doesn't make sense and logically one would think there is. I mean that 5 year old might be smarter than you, but you still won't believe that that he is, right? It's gonna take a while for him to prove himself. I think that let's, I mean, I think we're kind of kind of uh peaked here and maybe, maybe, maybe this is the right time to, to just go through last thought and, and then, and wrap up and we'll just come in, in this direction. Glenn, let's start with you. Yeah, thank you. So, so at AWS I'm honored and privileged that we've been thinking about this for a period of time. We established aerospace and satellite business unit at AWS 4.5 years ago. I was privileged to be asked to lead it as I retired from the US Air Force and US Space Force, and so we've been thinking about this problem for a while, to the point that even 2 years ago we did an on orbit. Demonstration TRL Level 9, where we built a purpose-built hardware software capability, put it on a satellite launched by our friends at DObit, and we ran a series of experiments for a few months on hyperspectral data where we determined if you can do on orbit pre-processing of data rather than having to push it to the Earth and process it. We reduced overall bandwidth requirements by 42% while achieving 100%. Accomplishment, right? So what it did is it allowed you to do more mission. It didn't save any money necessarily, but allowed you to do more mission for the same amount of expenditure. That's where we're going for the future. And at AWS, our team and in our new partnership with Ä¢¹½ÊÓƵ Allen, by the way, we just signed a strategic collaboration agreement together. Ä¢¹½ÊÓƵ Allen is now part of the AWS Gen AI Innovation Center, and so we're continuing to team on cloud. Computing edge capabilities and it won't surprise anybody if I tell you keep an eye on this space because the space team at Ä¢¹½ÊÓƵ and the space team at AWS along with a bunch of mission partners are continuing to push the envelope here on where we need to go in the next steps for on orbit edge computing and capabilities and so stand by more to follow. OK, so Omar, you're you're surrounded. So you get the next word. OK, so I think on edge will be eventually ecosystem of edges, and it's gonna be agentic systems where actually you have multiple things robotics, you know, non-robotic systems and IT IT components and agentic systems controlling and whole ecosystem of products that are on the edge. I think that's something we're starting to see right now and and it's gonna be much more amplified obviously in the future and we will be definitely levering on on capabilities like these for for longer missions as well. All right, Dan, you get to wrap it up. All right, I think, um. That's, that's a wonderful place to be. So, so one thing that one thing that I've noticed throughout, uh, throughout my entire career is that space technology has a funny history of being extremely useful here on the ground. And so as much as we're talking about sending agents to Mars, which is wonderful but might be a little hard for some people to understand. What is that going to do for us on the ground a lot, and this is why we do it. This is why we get up every day and say I'm going to look at the heavens. And my last thought here is is you know, just that, and I want to invite you all out to the lobby and check out our 4 demos. 2 of them are very relevant to Edge AI. We've got a RSO. we were able to not the whole mission but just the image collection translating RSOs are resident space objects. Can we see a satellites in a starry field. We were able to reduce that that data stream down by 94% in some of our studies didn't include any of the operational satellite systems, so that's. where a lot of, a lot of the, the extra, you know, 40% might might exist, as well as our, our technician in a box, a generative AI rag that fits inside of a box completely disconnected from the internet. When you log into Chat GPT, it works great. Turn yourself into airplane mode and you will watch it fail. This is what we're talking about. How do we make that work in airplane mode? So constraints drive innovation. That's what you guys, you specialize in constraints. That's all three of you. That's a Jeff Bezos quote. It's a good one. Go innovate. That's right. All right, thank you very much. Another great discussion. Thank you, Tobias, Clint, Omar, and Dan. Uh, as I think about, uh, talking about the edge, uh, you know, one specific project that we're working with NASA on, uh, Ä¢¹½ÊÓƵ Allen, is, you know, the lunar terrain vehicle, pressurized rover. Uh, in their attempts to go to Mars, but there's another, you know, edge device that, uh, it kinda encapsulates this whole day when I talked about actors, is the spacesuit itself, becoming an actor, right? The, the spacesuit is an edge device of itself, and I think that um being able to You know, think about that along with the discussions, um, you know, now the living thing that's working, uh, to support, uh, that astronaut or that actor, um, being in the spacesuit and the edge device, we, we've genuinely move the conversation forward, uh, and it's exciting to have this, this panel together.