really sit at the intersection of so many essential things right now. National strategy, innovation, AI, all of these are on your plate. Not to mention nuclear, what's going on with space, the list goes on and on, but we only have 20 minutes. So, I want to dive right in to what you see as your key priorities right now and how you measure your success.
Yeah, I mean our key priority across all of our different topics is ensuring American leadership in emerging technology. For us, we believe that emerging technologies are going to truly unlock great American prosperity. They're going to be the drivers of economic growth. They're going to be the places where Americans are going to be working and they're going to be bettering the lives of everyone across the country. So for us, we have to maintain our leadership in things like AI, quantum, nuclear, advanced manufacturing, space. And for each of those portfolios, our office helps coordinate kind of a national strategy across them. So what's very unique about the United States versus a lot of countries in the world is there isn't a tech or science agency or ministry, if you will. All of that work is spread out across a number of departments. So the White House helps coordinate that effort. There is important diplomatic work that's done, for example, at the State Department. There is research and development work that's done at the National Science Foundation or places like DARPA or our national labs. And the one place where we can bring all this together and make sure that all the pieces of government are talking to each other and moving ahead is essentially our office.
Yeah. And so you're sitting at the middle of all these different departments, and when it comes to this very complex question of AI, there's been a lot of attention recently on Anthropic. Anthropic doesn't want its AI to be used for unrestricted military use, and then the Pentagon responded by blacklisting Anthropic as a supply chain risk. So how are you advising the government to use Anthropic?
So broadly, the Anthropic question, especially around Mythos, kind of touches a wide variety of policy areas. So I think it's important to sort of break those down in different pieces and understand the policy process behind them. So the first, before Mythos was even released, was this question of what types of models should be available to American warfighters in classified environments. And for a very long time, the Department of War was actually reliant on a single vendor. And that's obviously not in the best interests of the department or of the warfighter itself. So, we spent the last two months trying to figure out how to solve that problem. And on Friday of last week, we announced that all the other major AI vendors are now on board for using their models in classified environments. And again, the north star is we want to provide the best possible tools to our warfighters as they're abroad doing the very hard and very challenging, very important work that all of us as American citizens rely on and actually appreciate. Now the second piece of that puzzle is this question of Mythos, which is a pre-release model, has not been released publicly. And the question around that is how do we use that model in order to ensure that we can harden government systems and set up a structure and a plan around how a lot of the other critical infrastructure and critical systems in the private sector are also protected going forward. And that's a more discrete cyber response question that is being run by our national cyber director.
So with Mythos, there's also this concern—Anthropic has said Mythos is too powerful to release publicly, to deploy safely, to be able to deploy it. What do you think of this strategy of a limited release for Anthropic? And how do you think about preparing the government for what cyber risks may come because of Mythos or other similarly powerful technologies?
Yeah, I think broadly speaking, what you want to do when these new powerful models come out is ensure that our most critical systems are protected. And what's very unique about Mythos is that it is very good at identifying cyber vulnerabilities. And these are issues that we have within our own government systems, which individual companies have disclosed publicly that Mythos has been able to identify these vulnerabilities, and in some ways we should think of this as an opportunity. We have a moment in time where we can sort of build the strongest cyber fortress in the world by using this model to identify these vulnerabilities, to patch them, and ultimately make them resilient to adversaries who may want to penetrate them.
How concerned are you about cyber risks right now?
I think generally speaking, we're at a moment where our sort of white hat actors have availability to the Mythos model as well as GPT-5.5, which came out, which has shown that it also has really good—and make sure we're pushing ahead with this necessity of a buildout for AI leadership, but also doing it in a responsible way where individual Americans are not negatively impacted.
Yeah, it's interesting thinking about your job and how you're working across all of these different departments and agencies, and the OSTP doesn't directly control budgets. So, you have to figure out how to drive alignment in what could be a very fragmented landscape. How do you prevent a fragmented system where different agencies have different rules and different policies or different standards about what types of technology or AI they feel comfortable working with?
Yeah. So, one bucket of work is around our research and development dollars. So the federal government spends about $200 billion a year on R&D, and that's spread out across a wide variety of agencies, from the National Science Foundation to our national labs, Department of Energy. And every year I write a memo with Russ Vought, who is the director of the Office of Management and Budget, where it essentially sets the presidential or the White House priorities for R&D, and then that document is then used to inform the budgets across those agencies. So you kind of use this—sounds very government boring, but there's memos out there that essentially direct agencies on what to prioritize, and that's a good way to kind of get at least the R&D dollar shaping together. And broadly, for a lot of our other technological domains, we end up sort of running or creating national strategies which give direction to individual agencies on actions they should take. And some of you may be familiar with the AI Action Plan, which the president released in July of last year. And that was kind of our national strategy as a government on how we can ensure that America wins the AI race. And it talked a lot about individual actions agencies can take to ensure that we're leading in innovation. So those were like regulatory-related issues or R&D issues. The second pillar was all about infrastructure. So back to what we were talking about, the ratepayer pledge: how do you make sure that we're building out the right infrastructure to power the AI revolution? And the third bucket is about international engagement: how do we make sure that ultimately the whole world is running on American AI?
And you were really an architect of the AI Action Plan. So this was something you worked on, and in doing so, I'm wondering how you think about balancing deregulation to enable AI innovation with, on the other hand, risk mitigation, because there are so many risks associated with AI.
Yeah. So generally, I think when we think about regulation, what's unique about the US versus a lot of the rest of the world, our north star on broadly on regulating AI is it has to be, at first principles, use case and sector specific. So attempting to create sort of one single law for AI ultimately kind of falls apart. The Europeans have attempted to do it; it hasn't gone very well. And when we say use case and sector specific, we mean that for individual use cases, the relevant agencies should be doing the regulation. So if you are creating an AI-powered medical diagnostic, the FDA should be the one regulating that. If you are creating a commercial drone, then the FAA should be regulating that. And kind of the list goes on. And in that environment, you can imagine that each of our existing regulatory agencies build capacity in order to regulate the AI applications that apply to their system. Now, on the same token, there are very important national frameworks that have to be in place for people to have more confidence in AI and for it to be ultimately able to proliferate more broadly. So, we released a draft legislative proposal to Congress last month, and this was our attempt to figure out or at least put forward to Congress: if you did want to pass an AI law, what would be in it? An example of that is protecting children. That's something that we as an administration deeply care about, most American parents care about. So how is it that we can create a system that actually mandates that technology companies provide tools to parents so that parents can make decisions on what their children see, who they interact with, and how the AI interacts with their child? So that's an example of the type of national framework we look at.
And so when you look at these different categories where pharmaceutical use of AI will be governed through those bodies, when you look at something entirely new like a chatbot or generative AI tools, where do those fall, or are those governed under the child safety category that you just mentioned?
So generally, if you apply a large language model to a health issue, for example, then the FDA would be the one that would be regulating that use case. And if you commit crimes using a chatbot, for example, then existing law on hand that regulates those particular criminal activities already exists. So, at first principles, you always have to see what law is already on the books and then start thinking about what else needs to be added given the chatbot. Now, on the chatbot front, I think children are very, very important to protect, and that's something that the president has spoken about, the first lady has been very vocal on, and we hope that Congress can ultimately move on.
And when it comes to this Anthropic issue, Mythos issue, it seems like that raises questions about dual-use AI for both consumers and also government access or military use. Do you think it would be useful to have a framework for approaching specifically these dual-use companies?
