From Action Plan to Mission Impact: Key Takeaways from America's AI Action Plan

In this inaugural episode of INSA’s All Source Podcast, From Action Plan to Mission Impact: Key Takeaways from America's AI Action Plan, hosts Chitra Sivanandam and Dr. Yevgeniy Sirotin break down what it takes to move America’s AI Action Plan from policy to execution. They examine the shift from experimental AI to operational capability, highlighting the need for scalable infrastructure, secure data, and stronger public-private collaboration to deliver results at mission speed.

The conversation sets the foundation for the series, outlining the key challenges and opportunities ahead as the national security community works to translate strategy into real-world impact.

Transcript

Chitra 0:02

Welcome to the Intelligence and National Security Alliance's All Source Podcast from Action Plan to Mission Impact. I'm Chitra Samanatham, president of Living Dino and chair of the INSA AI subcommittee. And as a quick background, what brings me here, what what brought me to INSA and to doing the subcommittee in general is I've been in the IC space for over 25 years and have always probably leaned more on the analytics side of things, which I think these days put the buzzwords in all analytics equals AI. But as we think about how we're deploying all these capabilities and how we're optimizing what we do and delivering for our customers, I mean, everything comes down to AI and everything's comes down to how do we improve our decision factors and improve our agility. So that's why I'm here. I'm I'm trying to figure out how do we do this smartly and how do we do this cooperatively with industry.

Yevgeniy 1:04

And I'm Yevgeny Sorotin, uh scientist director at SAIC and vice chair of INSA's AI subcommittee. And my background is a little bit different. Um, so it's going to be great to kind of share. Uh I'm a neuroscientist by training, and over the last uh 12 years, I've been transitioning to uh to working in evaluation of various types of AI systems, um, largely focused on topics of biometrics and identity and as they relate to national security. And why now? And I think now we're seeing really like an expansion of the capabilities of AI beyond just typical machine vision applications and maybe simple decision systems to more general systems that operate on principles that are similar to those uh that our own brains are based of, uh large-scale neural networks. And I'm very much interested in understanding how we work with these new non-deterministic systems while still being able to understand where they perform well and where uh what kind of risks they introduce. So I'm looking forward to discussion on during this podcast.

Chitra 2:12

Yeah, and I think what's gonna be fun is like I've done a lot with the startup community. So I'm kind of looking at how do we go, go, go. And uh, I'd love the temperament of like you saying, okay, well, let's make sure we're paying attention to the risks. So hopefully we can um figure out where we're going and and where we're uh uh slowing down and doing it all thoughtfully and figure out that healthy balance. And the underpinning of this and the impetus for putting this podcast together was really looking at the White House's uh America's AI Action Plan, which lays out a very ambitious vision. Uh and turning that vision into operational capability is really where the work begins.

Yevgeniy 2:51

Awesome. Yeah. No, over the course of this series, we're gonna be joined by uh a number of guests who are leaders across industry and academia. And we're gonna try to break down what it actually takes to deliver AI at mission speed uh from infrastructure to trust, workforce, and uh will address global competition as well.

Chitra 3:11

So I think on that note, let's just dive in.

Yevgeniy 3:14

Sure. Yeah. And I think let's start by asking the question, you know, why this series and and and why now? Chitra, uh, if you could start us off.

Chitra 3:23

Um over the years, I think it's uh no longer an anecdote that our attention spans have gotten shorter. And um, with more activity happening and more knowledge being easily accessible, I think we have to figure out how do we think about the right medium and the right format and the right way to engage across a bigger community. So A, that's why podcasts, um, now is the time for action. The uh the time for talking, the time for thinking, uh it's still it's still there, but uh the uh AI action plan has really called out more of a call to action. So we're trying to figure out how do we do that, how do we engage across the the community, across our government leadership, across the INSEM membership, and um no longer think of this as experimental related to AI, but really it's operational. So how do we start doing things operationally?

Yevgeniy 4:12

Aaron Powell Yeah. And as this uh technology you know progress is progressing so quickly, I mean, just over the last few years, you know, we sort of have to adapt to moving at that speed and maybe making adjustments at mission at the speed of the technical development rather than at the speed of maybe, you know, old, older uh uh procurement timelines and things like that.

Chitra 4:32

Aaron Powell Exactly, exactly. So time for action. Um and I think one of the key things is the plan as uh we read it, it's really centered around three pillars: accelerating innovation, building American AI infrastructure, and leading in international AI diplomacy and security. Um I think from that perspective, we are trying to uh tee up uh some some great uh leadership and is and speakers to weigh in on some of this. How do you think about those topics as it relates to the plan and where we're going?

Yevgeniy 5:05

Yeah, no, I think I think it it's important to us to understand, you know, what kind of AI infrastructure do we really need to build? Is it all about uh you know building? Of course, we understand we need to build these large-scale data centers and provide the energy required to operate these things at scale, but what else is there? You know, how do we uh ensure that we have the right data available for these systems? How do we make sure that the data is adequately protected, uh, but at the same time is available to maybe help fine-tune or or or or or um develop these systems? Um and asking the other questions about the infrastructure, how do we know they're working well? How do we make sure that we have the right governance in place? Aaron Powell Yeah.

Chitra 5:47

And I think hopefully we'll also get to like the soft, uh touchy-feely part of that ecosystem, too. How do we think about the people and uh those that are leaning into it and the skill sets that are required for the future and those that might be a little more tempered in their experimentation? Um I'm hoping that some of our speakers give us good anecdotes on how they they are thinking about it and what they've experienced and uh delivered to date.

Yevgeniy 6:11

Yeah.

Chitra 6:12

Um I think on that note, related to um kind of hands-on activities and and um experience. Uh what do you think about like the role of partnership? There's a lot of emphasis on partnership. How are we thinking about that?

Yevgeniy 6:26

Yeah, no, that's I mean, that's a great point. I think there's there's been this notion that I keep hearing again and again and again, this this notion of really successful public-private partnerships, because I don't think that you know any one group is going to be able to do all aspects of this. And we're gonna start trying to um, you know, trying to talk a little bit about what are the roles of the different partners in this, what are the roles of government, and what are the unique things that the that the public sector is gonna do? And what are the roles of uh of AI developers, but also other industry participants, like uh like integrators or evaluators of these systems. Um it'll be I think this I I'd love the speakers to talk through what that ecosystem looks like and you know what a success look like in the end um of a healthy ecosystem so that we can have the AI we need when we need it and also have that level of assurance that it's working well for us.

Chitra 7:19

Yeah, no, I like that. And you know, as we think about partnership, I think we kind of default to thinking that the government is risk averse and is not as uh adept at partnering with industry and and leaning forward on on new things. So um hopefully we'll have some really um good discussions with some folks who have both sat in government seats and have crossed over into industry seats and um understand their experience as it relates to AI and everything that's moving fast both from both perspectives. So we figure out really, you know, how did they accelerate things on the inside and how did they think about the barriers from the outside? I do think that's hopefully gonna be something that we end up being able to address. And I I for one think that um there is not gonna be a right answer. It's gonna be, I think, murky at best at the beginning, and we're gonna probably trip on ourselves quite a bit in these early days. Uh, but I think as we said earlier, it these are days of action. So we're just going to do and learn how to figure it out.

Yevgeniy 8:25

Yeah, no, absolutely. And and as part of that, we're gonna need to sort of start maybe asking these questions, you know, where should uh the government, um, which is driving this AI action plan forward, most lean heavily on industry. So what are those things, in your opinion, that you know, industry really needs to provide? Uh like is it the infrastructure build-out, but is it also um the aspects of responsible AI and just general coordination? Where do you think industry should play the bigger role here?

Chitra 8:57

Aaron Powell I mean, I think it's um it's all of that and significantly more knowledge, right? So I think industry has the advantage to experiment and um try things faster with um less reservations, but the good and bad of that is um maybe on the industry side we might do something without thinking it through from a security, privacy, and um safety implication. Um maybe there are more things we learned though quickly because we're able to do and offer that knowledge and um and guidance on the government side. Um my belief is that there's only so much we're gonna be able to do, having not experienced the good and bad on the inside, that some of the guardrail settings will happen probably more based on having kind of learned it the hard way from industry and being able being able to adopt and and transfer that knowledge to the government. And I think uh similarly on the government side, we're thinking about like why do we really care about this? What what is it we're trying to protect and really be thoughtful about it as opposed to um just lean on some rules?

Yevgeniy 10:05

Aaron Powell Yeah, no, I think I think the AI action plan certainly argues for, you know, let's let's deploy, let's, let's i implement. And I think that's a really great direction for us. But I think there's still going to be some aspects of, you know, this sort of little g governance that we're gonna need to sort of have in order to make sure that these systems are being deployed in a way that sort of lives up to the expectations that we have of them.

Chitra 10:30

Aaron Powell And probably like collective governance, right? Maybe that's the really the the theme under the partnership is we all have to have some sense of governance for each other because there's only so much we can test independently.

Yevgeniy 10:40

Aaron Powell Yes, yes.

Chitra 10:42

Um so uh on that note of implementation, like what do you think the guidance should be? Like how do you think about what it really means to move from like what we have been doing to date and what we think we might have figured out on paper and then actually moving it forward and implementing it?

Yevgeniy 10:56

Yeah, no, I think uh so sort of back back backing up a little bit. When we talk about AI, what are we really talking about? What's changed? Um and and and to my mind, like what's really changed over the last few years is we went from these systems that could you're really our experts at doing maybe one thing, like face recognition or object detection, uh, to now we're moving into AI systems that are capable of doing many things and are becoming generalists. And we talk about foundation models and you know, uh general intelligence systems. And I think that moving in uh in changing from that mindset of I've got a specialist system that I understand has got this superhuman performance in a particular domain, um, moving from that to an understanding that I've got this generalist system that maybe makes more mistakes, but has gr like this this really growing set of capabilities, I think that will be a barrier and a shift that we need to sort of think about as we reconceptualize what it actually means when we say AI. So I think that might be a a little bit of a of a change for us.

Chitra 12:02

Aaron Powell You've got me thinking, uh how how have you found yourself like working with AI and how how has that changed over the course of the last year?

Yevgeniy 12:12

Yeah, no, I I've become personally, I've become an early adopter of uh of AI. Um I use it to uh increase my productivity and and my company has been doing a great job of sort of making resources available to us uh in deploying these systems. But I would say over the, you know, I've been testing AI-based systems over the last decade. And when we started testing, that's big news was that we would get an AI system in biometrics, it was face recognition that really stood out that outperformed humans on this task. So we have a capability for face recognition. And in 2014, uh there was a deep learning model that for the first time showed uh performance on par. And very soon after that, it exceeded human-level performance on this task. So it was really exciting to see those changes. But as we get to over the last few years, people have discovered that actually when you start looking at foundation models, even when you apply them to these tasks that have been traditionally uh done by expert systems, they can in some cases surprisingly outperform those systems. So the one I'm thinking about is we've had a class of systems that I've worked with. They're called presentation attack systems. Uh, they detect whether or not you're wearing like a mask or doing something in front of the camera where you're trying to impersonate somebody. And it turned out, turn it's turning out that some foundation models are actually really great of doing at doing that task right out of the box. So I think from my conceptualization of how I approach AI has been from like, hey, I need to really large data set of a particular kind to develop an expert system that's going to do one thing, to now thinking about, oh gosh, these systems that are developed on this huge state, these huge data sets that are able to reason more completely about the world, uh, understand more about the data that they're working on, um, what else can they do for us? And sort of like, so for me over the last year, it's this sort of discovery that, hey, wait a minute, these systems can really start adding a lot of value in places. And the real question is, how do we do that at scale? Because certainly from a computational efficiency perspective, we're still not quite there yet.

Chitra 14:26

Yeah. And like I love the mask anecdote because that makes me think like, what I really love about everything we're seeing today is that it's it's all added the the adaptability is like beyond what I ever thought probably a year ago. And the mask example is a perfect one, right? So if I think about the way you were describing it, I remember those early days of trying to think about the mask recognition problem. And that was focused on um the terrorists, right? We're trying to figure out who might be hiding from the camera system. Versus then you look at the post-COVID shift and like the mask type algorithm capabilities was trying to figure out, hey, how do I do the better identification given that you're gonna be wearing a facial mask?

Yevgeniy 15:11

Correct.

Chitra 15:11

And so it was no it was a different kind of threat off the same type of algorithm with this a different kind of use case to give me a different type of response that became really interesting because now I'm looking at it for a different purpose. So it makes me think like really where we are today is it's this really good convergence of like I've got a lot of really highly adaptable capabilities and algorithms that I can tune to different missions so fast. Um, and that that becomes like an amazing challenge, right? And it becomes uh both from like the opportunity side and from the risk side, becomes a very interesting place to be.

Yevgeniy 15:48

Yeah. And how do how do we as as as humans, you know, how do we work with these systems? Because um because one of the paradoxical things that can happen with these AI systems is as they improve, um, and the human role of humans in overseeing these systems could become even more difficult. And I'm hoping that we'll have some guests that can talk about that that human element and how do we adapt our workforce in order to, you know, educate ourselves so that we can, you know, that we can correctly respond to uh these AI systems in in these different workflows that we might um, you know, implement.

Chitra 16:24

Yeah, I agree. And and and and there is like we can't be uh ignoring the fact that there is a pacing thread out there. So as we think about that, I'm hoping some of our speakers talk about our adversaries and and how we keep up in light of what we're seeing as those pacing threads. I think it's gonna be a fun podcast. But this was great. This I I'm really excited. This is a great way to kick off the series. I think that we've set the stage for some very important conversations, and I'm hoping that we get to it enlighten the uh the membership and the broader IC uh community writ large. In our next episode, Data or Bust, we'll be joined by Sean Batir to take a deeper look at one of the most critical elements of the AI action plan, data. And as the plan makes clear, we can't have world-class AI without world-class data.

Yevgeniy 17:15

And we'll dig into what it takes to treat data as a true national strategic asset from building high quality AI ready data sets to enabling secure access across environments and partners.

Chitra 17:29

And what it's going to take to turn that data into real operational advantage across the IC and the national security community.

Yevgeniy 17:36

Thank you for joining us.

Chitra 17:37

Thank you. We'll see you next time.