Host Intro (Gene Kim)

So let's continue against the theme that Joe described, the notion of embedding innovation closest to the edge where the problems are actually being solved.

So I learned this from Steven Spear. He said, "Yeah, the reason you need to do this is that leaders are too slow, too few, and too stupid." And it doesn't mean intellectually — it means that they're too far away from the problem to actually know how the problem should be solved.

And so I mentioned a workshop yesterday. In fact, that's where I met Matt Jones from Sigma Defense, where we got to train a bunch of people in the US Navy to learn how to vibe code. And the ultimate goal was to actually enable sailors to solve their own problems without being dependent on a busy and distant IT function.

And it was so amazing to see switched-on younger officers vibe code things. But what was really great was I got to meet Commander Collin Fox, who helps lead Task Force Hopper, and he has been working on these amazing innovation efforts across the US Navy surface fleet. And I thought these use cases were, on the one hand, very novel, because they're on boats and ships and so forth, but also familiar. And I thought what made it particularly interesting is that it represented massive distributed experimentation at scale, where it's not one central group responsible for coming up with the ideas and implementing those ideas.

I asked if he would be able to share those stories, and I was so grateful that he said yes.

Collin.

CDR Collin Fox

Hey, thank you so much, Gene.

I am Commander Collin Fox, Chief Strategy Officer for Task Force Hopper and representing US Naval Surface Forces.

I am so excited to be here, after honestly being pretty terrified at the initial invitation to speak. I was telling my family last night that it is so much more fun to be here than any defense summit that I've ever been to. It's not a high bar, but this is really fun.

So I've been in this role as CSO for just a little bit over four months. And the changes in tech and policy have just been absolutely immense. So the good news, for me at least, is that we've all been restarted at the same time. And so I'm a little bit closer to the starting line.

All right. So a little bit about me.

Briefly about the United States Navy. We face every challenge that a Fortune 100 company faces: legacy infrastructure modernization, talent competition, supply chain complexity, cybersecurity, obsolescence. We also have some added challenges of operating in austere environments while sometimes being shot at, only promoting from within, making systems that must work when disconnected from the cloud, and, last but definitely not least, the joy that is the Federal Acquisition System.

So, a little bit about me. I am a generalist. I'm a very curious generalist, and just by nature, and a leader by very slow maturation. Every single assignment in my 22 years in the Navy has been incredibly different. My dad was in the Navy too, and so the longest time that I've ever been in any one place was four years at the United States Naval Academy.

So I was commissioned in 2004 from there and served as a helicopter pilot for the first half of my career and then as a foreign area officer for the second most recent half — I think half diplomat, half naval officer. So in the past 10 years, I've been assigned to the Chilean Naval War College, the US Embassy to Panama, the State Department doing arms control, and most recently a major headquarters in Italy, the Commander Naval Forces Europe and Africa.

Just as an aside, I ran content in five languages for the African Maritime Forces Summit last June. It was nuts. It was in Mauritius, so it was also beautiful. I got pulled in to moderate a breakout session in French with the Francophone West African group. My French is not that good, so I am so happy that Gene is running this and not me, and also that all of this is in English. Mostly in English. There's a lot of words that I still don't understand.

So why have I pivoted to diplomacy with robots? Why am I here talking to you about technology? I do wonder that myself, but I also believe that every technology, back to the most ancient, is built for purpose, and it's for a human purpose, and the more radical the change in the technology, the more important it is to recompile the purpose from first principles.

So as I see it, and a lot of our other speakers have echoed this, this is less of a tech problem than a people and an organization problem.

Okay. So these observations are framed primarily around military applications, of course. But if you squint, they might apply to other cases too — maybe even your use cases.

So here's some big ideas.

One: Russia's ongoing full-scale invasion of Ukraine has lasted longer than the US participation in World War II. A long time. There are other wars and rumors of war and current events that I'm not going to talk about. So, bottom line, the end of history has not arrived, sadly. I'd rather be out of business. And it's not likely to do so anytime soon.

Two: just by simple correlation of forces, Russia should have been able to defeat Ukraine years and years ago. Ukraine's ability to rapidly innovate and rapidly adapt technology is a big part of why it remains a country and has not been annexed by Russia.

Three: changing significant bits is now many orders of magnitude easier than changing significant atoms, right? Code is free. And software can change the function of a 30-year-old, even 50-year-old platform. Just as an aside, the USS Lincoln was funded in 1982, laid down in 1984, commissioned in 1989 — before the fall of the Berlin Wall — and will probably decommission somewhere in the late 2030s, early 2040s. So think about that for your infrastructure and support planning or comparison. But we can do new things with an old ship with software.

Four: big organizations are much slower to turn than a 100,000-ton aircraft carrier. It takes time and effort and leadership from the top, but also initiative from the bottom. So my boss, a three-star admiral in charge of all Navy warships other than aircraft carriers and submarines, recognized these basic big ideas and formally chartered Task Force Hopper to move fast by applying artificial intelligence and machine learning to make the ships that we do have that much better.

So the Navy, as I said, has over 300,000 sailors. Task Force Hopper is growing to 15 people officially assigned, of whom I'm the first after many years of ad hoc staffing. So I like to think of us as a startup with an angel investor who's very generous, named Uncle Sam.

So that's where we started. Here's where we are now, building the plane as we are flying it. We're running pilot programs. We're onboarding new folks, so I am not alone — very lonely with just one person. No, there's others, but they're leaving me soon, and I don't want them to. And also creating a new detachment in Austin, Texas, called FleetWorks, to do organic software development from the bottom up.

So again, thank you, Uncle Sam. You're great.

Okay, so this is how we think about our pilot programs. The character of naval warfare is changing very quickly. Technology is changing even faster. And at the top level, there's this shifting frontier of possibilities between the two — between the task at hand and the potential solution. So discovering the art of the possible is not primarily a technical problem. It's a people and organizations and relationships and creativity problem.

So AI is an accelerant for these changes at every level of our hierarchies all at once, but it's also a very uneven accelerant. So it's disruptive by nature. Despite all this change, one constant is the need to really understand what we're trying to do — whether that's in air defense or software development or the entertainment industry. What's the use case, right?

So this is, think way back to Aristotelian philosophy. Aristotle would call this the fourth and final cause, the telos — the purpose for which a thing exists and is brought into being. So this is absolutely fundamental. If you are firing on the wrong target, it doesn't matter if you're more powerful or you put more effort into it, you're still firing on the wrong target. It will not solve your problem. So at the same time, we're also exploring these emerging tools that will help us get there, and they keep on changing every day, and my head is spinning.

So I'm going to be light on technical details as we go through our use cases and focus more on the why and what we're trying to solve for our use cases. I hope it provides food for thought, and maybe even sparks an idea for how you can deploy your tech — maybe even at sea, maybe even on our ships. It'd be great.

---

When we deploy hardware and software on a destroyer or any one of our ships, these are just some of the challenges we have. We can't just put everything in the cloud and point the API and say, "Yeah, connect," and do all that stuff. On-prem architecture is a must. But these ships — most of our ships were designed in the '80s. Real estate, cooling, and power are at a premium and mostly already spoken for. The data pipeline is a soda straw that can turn off in an instant.

The organizational challenges are especially significant and maybe harder to appreciate, because projected requirements for future systems tend to be moderate extrapolations from the status quo. There was nothing wrong, for example, with horse cavalry as horse cavalry. We all like horses. Horses work as horses. But designing a better saddle does not make a horse a tank, which is how some of our projections, you'd think that it was.

All right. Enough big picture framing. So these are the three use cases I'll summarize, and I'm just going to get right to it.

01Use Case 1: AI Sidekick for Radar Maintenance

So this is the Russian cruiser Moskva after being hit by two Ukrainian missiles. Open source intelligence analysis noted that the fire control radars are stowed, as you can see from the red arrows.

Radars are a bit like seat belts, right? They only work when you use them. But unlike seat belts, they are very fickle beasts. And they're hard to maintain over many decades. As we see from the Moskva, really bad things can happen when you don't use them or they don't work. Guided anti-ship missiles have been around since World War II, and they keep on getting better. All the more reason to make very sure that our radars work better than the Moskva's.

So that's our why, one of them: keep our ships floating, don't turn them into submarines.

So this is our radar. It's better than the Moskva's. For comparison, our air defense system is about 40 years old. The Moskva's was about 40 years old. But the SPY radar — that's the name of the radar, it's not an intelligence agent, it's just what it's called — the SPY radar has been upgraded in hardware and software over the years. And that creates its own challenges for obsolescence, for standardizing tech baselines, and for training.

So the initial complexity and the subsequent upgrades for a radar that can literally help shoot down ballistic missiles — hitting a bullet with a bullet in space — means that maintenance is going to be a challenge. Our most junior sailors go through almost a year of technical training before they can even touch the radar, and mastery obviously takes many years. Even so, it's a really time-intensive process to keep these systems up and running, often needing field technical reps to fly out and join the ship underway.

So with those challenges in mind, this is how my team used generative AI to get after the very real problems of SPY radar maintenance so that our ships don't become submarines. We worked with the various stakeholders inside the Navy to run a small pilot focused on improving the efficiency and outcomes of radar maintenance.

So from a technical perspective, this is not cutting edge, and you all should know that. It's a standalone laptop with a prosumer graphics card running a RAG. I think it's a 5090 with 24 gigs of RAM, and an open-source model — I think it's Mistral 7B. There's continuing refinement on the system instruction with the terms of art for the radar system itself as part of the prompt engineering. The front-end user interface is customized so that it's more than just Ctrl+F.

There are so many benefits as compared to running maintenance off of paper — especially, or even PDF manuals. I'm preaching to the choir here. There's automatic continuity of logging between the different maintainers. It can combine troubleshooting and corrective maintenance and preventative maintenance — when you're all looking at just one panel — into one streamlined decision tree.

Our early experience with the tool — it's still in early phases of this pilot — is that the main benefit is the time saved for the maintainers, and it's not getting significantly better maintenance results right now or more uptime for the radar. But, as I said, it's early. Oh, by the way, the ship on the right of my slide background — that's the USS Chafee, and that's the first ship that it's going to be deployed on, and it's firing an air defense missile. So it is appropriate.

02Use Case 2: DECC — Data Edge Collection Kit

Okay. So that was AI sidekick for radar maintenance, and now I'm going to talk about DECC. By the way, there's a lot of acronyms here, and they're all whimsical.

I'll start with story time about DECC. When I deployed on the USS Harry S. Truman back in 2007 as a helicopter pilot, from time to time you'd hear, "Away the Snoopy team, away! Contact starboard side." Photographers like this one with lenses as long as your arm would go up to Vulture's Row on the bridge wing and start snapping pictures of passing ships. And these pictures would eventually be uploaded to a database for future reference, for intelligence, for boarding, whatever. And each step of that process was touch-intensive. It was a long human process.

SNOOPI is one of the first of these ridiculous acronyms. It's the Ship's Nautical Or Otherwise Photographic Interpretation and Exploitation. Somebody got a medal out of that, I'm sure. Or otherwise. So as a helicopter pilot, I'd fly around and my crew would do the same thing with more humble cameras — like this guy in the middle window there. Take pictures of ships, other items of interest at sea, debrief them for processing.

So here's a clear use case ripe for AI. Use every platform available to collect data about your environment. The constraint wasn't data collection — it's easy to take a lot of photos, many people know that. It was data processing, storing, aggregation, and indexing. It was touch-intensive. AI has greatly expanded both our ability to process data and the applications for that processed data.

So think about how many hours have been logged with a driver behind the wheel before Waymo or any of the other companies in that space went fully autonomous. Now transpose that to the nautical domain. The Navy's unmanned surface vessels need data for training just like Waymo taxis need data for training. So we're putting these data edge collection kits on manned ships to automate and categorize data that was being collected and routed manually — for navigation and for anything else that we can wring out of that data.

So the Secretary of the Navy, who has a speechwriter and I don't, highlighted this in his speech a few months ago. He said that senior leaders really understand: if you don't build a data engine, you will not build an AI-enabled force. So that's why the Navy is deploying DECC.

"DECC allows ships to collect operational data at the edge and retrain models. It turns ships into learning systems, not static platforms, enabling an iterative and adaptable feedback loop with legacy bespoke architectures that historically evolved only through programmatic redesign. This is how we move from demonstration to dominance." End quote.

Okay. So what is DECC? I've talked about the why. I wish I had a picture and the detailed specs — I don't. The SECNAV just talked about it in an open space two months ago, so I'm glad I get to talk about it at all.

It's an air-gapped system with electro-optical cameras — think FLIR turrets — feeding image recognition hardware at the edge. So instead of trying to offload terabytes of maritime dashboard camera footage through low-bandwidth connections, SATCOM, the system churns through that data at the edge and picks out the items of interest. The bandwidth demands go down by orders of magnitude. And obviously we all know you need real-world data to train AI for the real world, or at least it helps a lot. That data needs to be representative and more or less automatic, and not just humming along in the background.

03Use Case 3: FleetWorks — Upskilling Sailors for Software Development

Okay. So let's step back briefly for a moment from radar maintenance and automated data collection and think about similar tools for everything else that happens on a warship.

The network and development architecture reflects historical organizational architecture, for better or for worse. These are expressions of Conway's Law at sea. The organization responsible for the ship's networks has created a different on-prem RAG implementation, also standalone, per IT report. It's called the Comprehensive Help Desk Information Expert for the Fleet, or CHIEF. Another wonderful backronym. And here's another implementation from another organization — same thing, different system — called Systems Command Acquisition and Integrated Logistics Online Repository: SAILOR. Yeah.

So cute acronyms aside, there could be three or more RAG implementations with different models running on different hardware all on one ship with 300-something people on it, and that could just be the tip of the iceberg.

So when I gave a version of this brief at the Marine Corps AI conference a few weeks ago, some Navy doctors came up to me after the fact and they wanted to do yet another standalone implementation for the independent duty corpsman — the medical professionals on these destroyers.

Innovation at speed is inherently chaotic and sometimes duplicative. That's what I take away from this. There are going to be branches that diverge and later converge or are pruned. That tension, that friction is just part of the process. My long-term vision is that we can containerize these implementations, put them on a VM or something — I'm not the tech guy here — and put them on networked hardware onboard the ship like an on-prem cloud.

This is the third and final topic, and I'll be brief on this. It's upskilling a small group of sailors to help come up with and run the next generation of software pilots.

So let's revisit some of the big ideas behind our why. The rapid technological adoption can create significant advantage in conflict, and software can be updated for effect faster than hardware. Our adversaries are not stupid. These same facts are evident to them. They live in the same world. They are also trying to move fast.

So this is a race. The closer we can get to the right people with the right skills in the right place and the right resources and authorities to the point of need, the faster we can act — and the more likely it is that we can outpace those who would do us harm — and underscoring deterrence by denial, because honestly, I would much rather not go to war. It's better for everybody if everybody says, "Yeah, not today."

So this final example is a pilot of pilots that we are still in the process of fielding. We call it FleetWorks, and it's part of our larger digital workforce development program. So honestly, I'm just copying the Marines here. They've had a great ROI on their software factory. We've recruited sailors who have already got experience in this, and we're running them through a short accelerator course in Austin, and then we're going to let them loose on the fleet.

04Closing: How You Can Help

So here's the help I need.

There are lots of problems that we in the Navy are trying to solve — putting AI at the edge, dealing with constrained space, weight, and power. If you have similar constraints, I'd love to talk to you.

Second, there are a lot of rock stars in this room, and if you would want to talk to our FleetWorks folks that we're going to stand up in August, I would love to connect you with them. That would be great. It's going to be a small pilot.

And finally, if you want to sell to us or deploy your tech to sea or just talk, you can email me or use the QR code. There's a market research survey there.

And I'm over time. Thank you all very much. This has been really great. Thank you.