3 Takeaways - The Hidden Weakness in America's AI Lead (#315)

Episode Date: August 18, 2026

Could America build the world’s best AI and still lose the AI race?What if the biggest threat to America’s lead isn’t China’s technology, but our own fear of using it?Katrina Mulligan is Head of National Security Partnerships at OpenAI and previously served in senior roles at the Department of Defense, National Security Council, and Department of Justice. Few people have seen both Washington and frontier AI development from the inside.She argues that AI is no longer a technology question for the future. It is becoming core infrastructure for economic strength, national security, and the way we work.She discusses:●  Why America could win the race to AGI but lose on adoption●  Why trust in AI is dramatically higher in China and the developing world●  How China and the U.S. are taking fundamentally different approaches to AI●  Why AI may ultimately transform society as profoundly as electricity●  The biggest mistake leaders make when trying to adopt AI●  Why AI transformation cannot simply be delegated to the IT department●  How she has used AI to become at least 30% more effective at her own job●  Why using AI like a search engine means dramatically underestimating what it can doAmerica may currently have the advantage in building the most advanced AI.Katrina explains in today's 3 Takeaways™ conversation why that lead is far from guaranteed, how fear could become a strategic disadvantage, and why the countries and organizations that learn to redesign themselves around AI may ultimately have the most to gain.

What if we're asking the wrong question about AI? Everyone is focused on who will build the most powerful technology. But what if the real race isn't about building AI at all? Hi, everyone. I'm Lynn Toman and this is three takeaways. On three takeaways, I talk with some of the world's best thinkers, business leaders, writers, writers, politicians, newsmakers and scientists. Each episode ends with three key takeaways to help us understand the world and maybe even ourselves a little better. Today I'm excited to be with Katrina Mulligan. Katrina is head of National Security Partnerships at OpenAI, the company that created ChatGPT.

She previously served in senior roles at the Department of Defense, National Security Council, and Department of justice. She was number two in the Pentagon overseeing special operations. Few people have seen both Washington and Frontier AI development from the inside. She joins us to discuss why AI is unlike any technology revolution before it and how it will transform the way we work, compete, and solve problems. Welcome, Katrina, and thanks so much for joining three takeaways today. excited to be here. My pleasure. Katrina, you have a great quote. You said something like the world is looking at a calendar while Open AI is looking at a watch. What does that mean? I'll start by saying when I made the move from the Pentagon to Open AI, it was the hardest professional transition in my career.

And I think part of why it was so challenging is because I wasn't expecting it to be so hard, because I felt like I had had really difficult transitions before, but there was something really different in destabilizing about moving from probably the most hierarchical and bureaucratic part of the most hierarchical and bureaucratic organization maybe on the planet to the exact opposite of it in every way. One of the other disorienting things about the transition was that people measure time differently at Open AI and time passes differently at Open AI than anywhere I've ever been. And we often say internally that a month that Open AI feels like a quarter anywhere else.

And so once you've hit three months in or four months in, it's like you've been a year anywhere else. And I try to help government understand that by saying frequently that like you're looking at a calendar and we're looking at a watch. It is an indication that the unit of time that we are thinking about that we have ahead of us needs a complete recalibration. And that's because the technology is changing so quickly. as well as the opportunities?

It is, and it's because, you know, we've long anticipated that eventually the models would get good enough that they would begin to conduct the research that improves themselves. And, you know, this idea, which is known in the AI industry as recursive self-improvement, is not quite upon us yet, but it's getting really close. And so, like, I can't even think five years into the future in terms of where this technology will be by then because it's so impossible. It's so nonlinear at this point. we're really reaching much more of an exponential.

What are other ways that make this AI revolution fundamentally different from every tech revolution before it? The biggest one from where I sit is that this is really the first time in American history that a technology of this consequence is being developed exclusively by the private sector with no government involvement. You think back to the advent of electricity, nuclear fusion, the internet, the human genome project, GPS, Yes, all of those are examples of technologies that were developed with government as a stakeholder in the development of the technology. They had a seat at the frontier table. They may not have been driving the whole thing, but they had at least a finger on the steering wheel.

That is not true about AI. And it's really unique in that all the AI labs are operating really independently of government and government doesn't have a major AI initiative that can in any way compete with or compare with what is happening in private industry. And I think that's not necessarily a bad thing or a cause to be concerned, but it's been a bit uncomfortable for government and destabilizing for how we normally think about how government involvement in technological advancements should proceed. And I think it gives the government maybe more blind spots or at least different blind spots than the government would have if they were involved in the creation or the genesis or

were somehow a direct stakeholder in the development of the technology. Most people think AI is chat GPT. What's the much bigger story that they're missing? Well, let me start by saying chat GPT is pretty awesome. Chat is kind of the entry, it's the gateway. What it did that is phenomenal is it took the barrier almost entirely away, the technical barrier to regular people of all kinds, being able to access and use this technology and experience its benefits.

You don't have to be a developer or you don't have to know anything about technology. But for enterprises and particularly for government and also for the future, if you're like envisioning like where is this all headed, it is really workflow transformation that is going to be the central thing. And it is the idea that every organization, whether it's a government or an enterprise, has some number of core workflows that are central to what they do. If you're a drug manufacturer, it's like the process of evaluating and testing different types of pharmaceuticals. If you're the intelligence community, the analytic workflow and the collection

workflows are good examples. Every workflow you can imagine has the potential to be transformed using AI. And it doesn't mean replace. Notice that I'm not using replacement terminology because there are parts of those workflows that can be enhanced or changed or maybe made more efficient using AI. but there are also in every single one of those workflows, parts that can't be. When I look ahead, I see chat, particularly for government, as being kind of necessary but insufficient, I think where government in particular needs to go, we aren't there yet, but we're on the path,

is really starting to reimagine what might be possible if we, you know, thought about our core workflows in a different way. Two years ago, AI seemed impressive but limited it. Today, it reasons what changed. When ChatGPT was first rolled out to the world, fun fact, there were only 200 people at Open AI at that time. There was a belief in the company. We almost didn't even roll it out. It was considered a low-key research preview. And, you know, the people who had built Chat, they thought it was like really interesting and impressive, but it wasn't something that we thought

was going to be come the moment that it became. And most of us internally, you know, if you look back at what that initial chat model, which I think was chat GBT 3.5, it might have been 3.0. But if you look back at what it was capable of doing, I would describe the outputs as like junior high school level outputs. It was interesting because it could like talk back to you and it was a computer. But it wasn't all that impressive in terms of actual raw intelligence power. In fact, I think any of us, if we went back and tried to use that old model, be wildly underimpressed with what it was able to do in contrast to how shocking the growth moment was at the time. What happened is really two things. One is that Open AI had had the sort of first fundamental discovery in the AI space,

which was the scaling laws, the idea that you could take these large language models. And if you increase the amount of compute by 10 times and the amount of data by 10 times, you could predict, predictably better outcomes in terms of the performance of those models and the intelligence capacity of those models. That was the original discovery and the roadmap, and it has remained true over time. Remarkably, I think for a long time, there was a belief that the curve would eventually flatten, that eventually you'd get to a level of capability where, no matter how much more compute or data you added, you couldn't actually continue to see linear progress in the same way. But that is actually not borne out in the data so far, and it doesn't show any signs of

changing. But there's another thing that happened in the same time window that wasn't so much of a research breakthrough necessarily on AI, but dramatically changed the way that we can use these models, which is we taught them how to use tools. We taught them how to use the kinds of the agenic harnesses that we've developed that allow the models to do search, to be connected to our emails and our internal document repositories and to navigate things on our computer and do computer use, once you have those abilities and you tie them together with the intelligence jumps and the reasoning capabilities of the models, now that's why it feels like in the last year or even in just in the last six months, it feels like something is different. And that's why,

because all of those things have now come together on the same runway and the same products. And they're now generally available and people are starting to realize they can do so much more than what they previously thought was possible using a simple chat interface. Most companies use AI to save time. Where's the much bigger opportunity? Time is one of our most valuable resources. And so just saving time is actually like a pretty valuable thing. We rolled out chat GPT across the state of Pennsylvania. They're workforce and we were able to measure that we were saving the average public servant eight hours a week. When you think about that, that is a pretty significant amount of saving. So point A is I think

that's a valuable thing. What I really am excited for and where I think hopefully public attitudes about AI in this country will shift when we start to see more of this is really around health, education, and scientific discovery. I'll give you an example of a thing. that really blew my mind. This was something we announced maybe a few weeks ago, still fairly new. Our company did a partnership with a children's hospital. And we often find ourselves, this is the same question I ask government. Like, what's a problem I could solve? That if I could just solve it, it would dramatically change things and you don't have time to solve. And like, where's that sweet spot of the kind of problem where AI is like uniquely helpful?

And the answer in this case was our undiagnosable patients at every children's hospital, there's some number of patients that have been in the system and been searching for answers for years often with no answer to be found. And if you think about like children and their families having to navigate the medical system in search of answers for really rare or really tricky chronic illnesses, and then you think about what is that? That is a data problem. It's a data and it's an analysis problem. You probably in every one of those cases are going to have images. You're going to have scans, you're going to have clinician notes, you're going to have data that spans, you know, multiple years. And you're also going to have alongside that scientific discoveries that are happening

independent of what's happening with that case. And so what we were able to do is get the right approvals and work the policy issues to be able to run our frontier reasoning models against these unsolved pediatric cases. And the models were able to suggest and propose new pathways, new tests that should be run, new potential diagnoses for clinicians to explore. And the net result of that were entirely new diagnoses for families that had never had them before. Now, one of my dream projects is to do that for the Defense Health Agency. What would be more beneficial than being able to help find new pathways to understand how to treat service members and their families in a better, more efficient way,

or to find diagnoses for people who haven't had them. Those doctors are able now to target their efforts toward a solution rather than, you know, revisiting in some cases decades old data to try to find a needle in the haystack. And it's a really good example of what is possible when you start being willing to creatively explore rather than being a defensive crouch. That sounds wonderful. When you look at organizations that are successfully using AI, what are the most important? important things that they are doing?

We have a lot of experience at OpenAI implementing AI transformations. And so we have a lot of insight and data about what works and what really makes a difference. And it has been consistently true that the biggest predictor of whether an organization will have a successful AI transformation or not is the extent to which their C-suite uses it. The people at the very top of the organization, it's whether they themselves are using the technology in their own work. And that is something that's important for government to really take on board because it's not enough to just have a bottom up approach where you just give everybody access to the tool and hope for the best. I always say you don't get fit by reading about working out and you do not get good at understanding how AI is going to transform your organization by reading about it or going to a panel discussion about it.

You have to be using it and getting in reps and sets yourself. And what's the biggest mistake that you see leaders making when they try to adopt AI? Probably the biggest one is building model gardens. And what I mean by that is the belief that the thing that you really need is just access to the maximum number of tools in a single place so that people can have access to 12 or 20 different models at their fingertips and that that's what success looks like. That, to me, is a mistake for a number of reasons. One, it's really expensive. You're buying every model and you're making it all available. Number two, Brenda from HR does not need 12 choices when she's typing a query into a chatbot.

That does not actually materially benefit. And in fact, I think it creates a lot of friction in terms of how the general population uses AI. I do think that model choice is really important. But for basic usage of chats, I do think that model gardens are something that we will look back on and think that was probably not a good thing for us to have spent that much effort on. America and China are taking fundamentally different approaches to AI. What are the differences and what are the implications? I think there are a few big differences. Number one, trust in AI is just wildly different, and that opens up a very different surface area. for China to explore and to create public support for things that it's doing that I don't see here in the United States. China is also much more heavily involved in the development of AI

than the U.S. government is. And some of that is a feature of the way that their society operates very differently from ours. The CCP and the Chinese government in particular has more of a direct seat at the frontier AI table than the U.S. government currently does. Although we are starting to see the U.S. assert its prerogatives in different ways, mostly around model safety and model releases. But I still think that fundamentally there is a different relationship that China has with the AI industry than the relationship that the U.S. has. In some ways, for the better, in some ways for the worst. I also think that the Chinese government and the CCP have moved faster to recognize AI as fundamentally an infrastructure play and to make strategic investment.

in infrastructure as part of their strategy. That is happening in the U.S. AI ecosystem, but it's happening because the companies are making those investments and because private industry is making those investments, not because the government has decided it wants to make them big play in that area. Could America build the world's best AI and still lose? Yes. In a word, yes. And it goes back to the trust deficit with AI.

I worry that we could win the race to AGI on the technology side, lose it on adoption because we are seeing much different approaches in the developing world to how quickly and how open those societies are to integrating this technology into what they do. And so it's great that we are at the leading edge of model capability and model development in the U.S. And I do think that if anything, we've actually strengthened that lead over the last year. But that lead is far from guaranteed. And we started by talking about the fact that AI labs look at a watch, not a calendar, when we measure time.

I think we've opened up the lead by a single digit number of months. Maybe we're six or eight months ahead of them now, whereas we used to be more like four or six. But we're not that far ahead. And that lead is not a guarantee. And so I do think that China has some really impressive and talented. engineers and they are really determined to win here. And I think we kind of have to approach it with the same seriousness that they are. What's changed in national security because of AI? And what do you see changing next? To be honest with you, I feel like we're just at the beginning.

I mean, one thing that's changed is I do feel that between last year and this year, there's a decided shift in who is paying attention. It's no longer just the chief AI officers. It's the mission owners. It's the heads of departments and agencies. We used to have to go knocking on their door to say, hey, there's a thing happening over here you should pay attention to. And now they are coming to us. And so it does feel like who is paying attention to this has changed for the better in a good way. What changes next is a matter that I very much hope to influence in that one of the The things I want to see change next is the level of resources and the type of resources that are being devoted to doing the kinds of enterprise transformation we talked about.

I want the government to stop funding AI like an IT item or at least stop only funding it, like an IT line item, and begin to start thinking about and allocating funding to real sizable programs of record that can get after really meaty, substantive mission problems. I would love for the government to fund a new way of thinking about the front door for citizen services. Like if you think about every single one of us, one of the things that unites us as Americans is that we have all had some god-awful experience dealing with citizen services and with some part of our government. And when you think about what a really crummy week looks like, it often includes like a trip to the DMB or having to prepare your taxes or some other way that you have to. to engage with your government. And that experience is often leaves something to be desired. That can be true at the state level, at the federal level, at the municipal level. And this technology,

look, we're not going to radically transform how much money we're putting it into those services. Like if we want it to be better, we're going to have to do more with less. And one way that you can do more with less is by actually harnessing the power of AI to transform the way that people are engaging with their government. If you could have that be even a slightly more positive experience in the aggregate, if each of those experiences was slightly more positive, over the course of a year, your experience with your government would feel really different. That's an indication of the level of ambition that I have for what government should be thinking about. That would be great when that happens. There's a lot of anxiety around AI. What gives you optimism? My own

personal experience with it, it's easy to be scared of a thing that you're not really incorporating into your life or taking advantage of. And it is much harder to be scared of a thing that is actually making your day better and that where you are actually experiencing how much more you're capable of accomplishing. I honestly think that I am at least 30% more effective, maybe more than I was a year ago because of how I've matured my use of these tools. I'm a better leader. I'm a better manager. I deliver more for open AI. And I just find that my own experience with it brings me a lot of optimism. Katrina, what are the three takeaways you would like to leave the audience with today? Number one, AI is no longer a future policy question.

It is a core infrastructure for public service and national power. Number two, trust is the rate limer. We can't just continue to be in a defensive crouch. We really need to think about the trust question. And number three is the real opportunity is redesigning work. It's not adding a chatbot to old processes. Thank you, Katrina. Thank you for joining three takeaways today. And thank you for your work to make AI safer

and to work with governments to try to make it safer for them and for all of us. Thank you. It's a great joining you today. If you're enjoying the podcast, and I really hope you are, please review us on Apple Podcasts or Spotify or wherever you get your podcasts. It really helps get the word out. If you're interested, you can also sign up for the Three Takeaways newsletter at Three Takeaways.com, where you can also listen to previous episodes. You can also follow us on LinkedIn, X, Instagram, and Facebook. I'm Lynn Toman, and this is Three Takeaways.

Thanks for listening.