So, we'll start with the obvious, the midterms. Democrats won control of the House and the Senate. The Democrats had already rejected the administration's proposed defense budget for '27. How might the election outcome impact the Pentagon's work, and are there any spending compromises you may have to make?
First of all, sad face. But the way these budgets always work is they typically pass the main defense budget before December. So there's still a chance in this lame duck session, they call it, where we hold both houses, that it could pass. But in the long term, the more strategic question is, yeah, this is sort of what happens when the parties in Congress and the president are different. The priorities shift, right? The other party is more likely going to want more domestic spending, less military spending. So that is a concern, obviously, for us at the War Department, because what we've been trying to do is make this transition, and we're at a transition moment at the department where we had all these weapons and planes and aircraft carriers that we built in the '70s, '80s, '90s to fight the Cold War, and they're old now. They take a lot of maintenance, they don't work as well as we want them to, they're not sort of recognizing the new threat with a Chinese adversary rather than a Soviet adversary. And what we've learned in Ukraine, Russia, is we need more autonomous systems. We need more drones. We need more massive, low-cost drones.
And so we're trying to make that transition this year and next year with enough budget that you can keep the old alive while you build the new. And that's what I worry about most.
How do you restack the blocks? What levers can you pull in the meantime to make sure that your utmost priorities proceed, and what are those?
You just have to make harder choices, right? So if you don't get the budget you want, just like in any business or any household, you have to make harder trade-offs. So the harder trade-offs mean what do you put at risk? What system do you say we're not going to do that because we have to build the new stuff? And those trade-offs just get harder when you have a different view of what the US military should be than what we have.
What's your number one priority?
My number one priority has been and is using super intelligence in the department for the enterprise use cases like you use in any business, for intelligence use cases, because we collect a lot of intelligence and synthesizing that, we're limited by humans analyzing things. Instead, I'd like to keep the same number of analysts but cover more ground. And then for war fighting, we use it because you want to see where all your assets are, how to do logistics, how to do all the things that prepare you to be more effective at battle at speed. So I've been doing that for about a year. We've had a lot of success, but that requires, guess what? Compute that we have to budget for. So that's the number one priority. The second priority is autonomous systems. You know what we've learned in Ukraine, Russia is robot-on-robot warfare on the front lines, and they're innovating at incredible speed on both sides there. And in the last administration, I think we built a thousand drones. So now we're at least up to, you know, we're trending toward half a million. I want to make sure that we have, you know, as many as we need. And they've done some amazing things for us already. In the Strait of Hormuz, we had the first autonomous boat rescue to down Apache helicopters. It was an amazing thing to watch. So I want to make sure we're prepared for that because I think that makes our war fighters safer.
And speaking of autonomy and warfare, as AI models become more sophisticated and integrated into defense, what are the administration's absolute red lines on how autonomy can be used and AI systems making lethal decisions on the battlefield?
Yeah. So the principle is, and I had this big dispute with one of these companies a while back, but that... Okay. The principle that we have, and it's written into our policies and it's been there for years, is that we have to have appropriate human judgment at every level, period. Because there are things you want systems to do that a human can't do. And I always give this example, in fact, I gave it to this company. It's like, okay, a Chinese hypersonic missile comes up. It's 13 minutes till it lands in the United States. You have a space asset there that can shoot it down. There's no guy in the joystick in the back room going, 'Do I got it?' You know, you're looking for a system that could do automatic target recognition to take it down. That's appropriate in that situation. So people underestimate the bureaucracy that we go through to be as safe as possible with any technology we integrate, but human judgment is the key.
Do you worry? I mean, I think one of the things that happens with autonomous systems is it boop-bops something out and there's a human inclination to trust it. It's an automated system that gives an answer and there's a certain confidence that humans assume in that. Do you worry about an overreliance on automated systems where that human judgment maybe doesn't show up in the same way?
I think if you want to talk about great powers in the world, China, you could argue Russia is still an important military power. We are the most conservative on those kinds of questions and always have been. And in fact, one could argue that there's a lot more... I was at Uber. I remember when we were fighting against the taxis, they're like, 'An Uber can't be safe. You have this driver. He's not licensed in the same way.' And then now people are saying, 'Well, a driverless car is not as safe as an Uber.' But if you look at the statistics, each one of those generations is safer. And I think you'll get to that way in autonomy, too. Autonomous systems, in many cases over time as the technology improves, will be safer than a human operator.
And I mean the other challenge of some of these autonomous systems, I'm thinking of Maven Smart System and others, is it's not just the AI models underneath it. It's the quality of the data and it's what happens when all of that comes together. And so given incidents such as the bombing of the girls' school in Iran, do you have confidence in the targeting capabilities of the systems that you have?
Yes. So that incident's under investigation still, so I won't talk to that, but I will say that targets are developed in a very sophisticated way at the department. They don't really use AI at all. What they do do is say, here's your assets, here's the enemy's assets, here's intelligence reports, here's satellite imagery, let me combine that for you, the human operator, so you can make the judgment. And I think that is the right way to use AI for us. But we're not there. We're not that sophisticated in doing these kinds of things that you're talking about just yet.
Let's talk Super Intelligence Force for a minute. What have the early weeks of that work looked like?
So it was a pretty extraordinary moment, I would say. So the president got... and you can imagine all these people in one room together, right? Elon Musk, Greg Brockman from OpenAI, Dario from Anthropic, Jensen, Sundar, the Microsoft CEO, I mean, all in one room to talk about what are we going to do about super intelligence to ensure that the American people feel like we're doing right by them. And it's working for them, not just for you companies. And I do think these companies have a lot to learn about how to market what they're doing. Just take the data center thing. There's a, you know, how do you communicate that this is a good thing for them.
Well, first you have to tell them that some of the rumors are not true about water usage, and maybe if you're using electricity, you have to dividend out, you know, things, and maybe you have to sponsor the local baseball team. Like, how do you become a member of the community? Saying AI has a PR problem.
It has a PR problem. And so we talked to them about, first of all, focusing on that, but then second of all, what are the standards by which you're going to hold yourselves to from a safety standpoint. And so we agreed on four-point principles about how we monitor safety. They're going to have third-party auditors, an independent committee of the board, so that anyone who has a problem with their safety practices, like the Hugging Face incident, they can go right to the board of directors. And so those are principles. And then I'm out here after this tomorrow and the next day I'm going to go talk about incident responses. How do they notify us when an incident happens? How do we categorize that incident? So what we're trying to do is create a framework that makes the American people have confidence these companies are going to do the right thing. Or else I think what you get is 535 legislators writing a piece of regulation that could kill the industry.
But does that group have teeth? Like, what can it do if you see something that doesn't... like, does it have the power to pause model development? Does it have the power to fine a company? Like, what can it actually do?
Yeah. So we're also going to lay out in the next week or two all the laws and regulations that apply to these companies. They already experienced what's an export control in two hours of not fixing a jailbreak. So that means you know you can't sell your product abroad at all and no foreign national could touch it or it's a criminal offense. So there are tools the government has on the books. Because the FTC, the Federal Trade Commission, has subpoenaed all these companies to ensure that their trade practices are fair. That's just the beginning. So there's a lot of teeth that we have to enforce some of the things we're talking about.
I want to talk a little bit about cyber. Obviously there's been a raft of incidents. We were talking about Buggeddon earlier this year, the PHP vulnerability a week ago. It feels like they're coming in sort of thick and fast. What is the Pentagon doing about the novel cyber risks posed by AI?
Yeah. Well, the bad news about these AI models is they find vulnerabilities that a bad actor can use to hack. The good news is if you're a coder yourself, you're at a company, they can help you find vulnerabilities in your own code and you could fix it. But I do think we're entering into a world where problems in code are no longer going to be sort of risk-free. So what we're doing at the department is not only patching our own systems, of course, but we're working at the Super Intelligence Force to patch critical infrastructure, to have the model companies help create patching tools for these companies, and to make sure that we have a better coordinated ability to tell the country that here's a vulnerability like PHP that exists, here's what you do about it and when. So we're trying to coordinate that. It's never been done before. So we're inventing that as we go.
How vulnerable is America's critical infrastructure?
I mean, I think that if you want to compare us to other countries, I think we're the least vulnerable big country because our hardware is the newest. We do the most, we have the most software that we produce. We're always on top of these things. But it's still, you know, imperfect.
Mhm. Yeah, speaking of weak spots, let's talk about the race with China. And some of the efforts that you've been interested in in sort of shoring up what America can do at home. First of all, in framing the race, how far behind the US do you view China at this moment?
For super intelligence, I view them six to 12 months behind. And you know, I hope it stays that way. We have superior chips, superior networking within a data center. We have something like 70% of the world's data center capacity, which is an important ingredient here. We have some of the best science and math people, or physics and applied math, which tend to be the most important people at these labs. China has a lot of science and math people, don't get me wrong, but...
...and models that have really come in and gained a following. Yeah. Did that surprise you?
No, because they're cheaper. Right. So a Chinese open-source model, open-weights model, is a lot cheaper. So you've seen the percentage of use move toward these Chinese models because companies, mostly small and mid-size companies, are saying, well, why am I paying this when I can pay this? So what I've been encouraging is for American companies who do open source to compete with the Chinese models. And there are some good signs on that with Nvidia's NeMo, Mistral, Reflection AI.
Why not ban those models? Venode Kos was here last night. He's advocated for more stringent regulation of open source. Why not take that step?
So I think banning something where you're going to force a price increase on the consumer that's maybe 50x, where the alternative is in a few months, not years, you can have American open source be equivalent, then I think the trade-off for that consumer or that business is more clear.
Where do you put the risk of a Chinese invasion next year? How are you thinking about the threat that China poses Taiwan?
You know, I'm not in that part of the business, so to speak, but I think, you know, I want to make my job by building all this new technology and the super intelligence is to make sure that every day when President G wakes up, he's like, 'Not today.' So I hope it's not '27.
Me too. All right. So let's come back home to Anthropic, which you alluded to earlier. You've declared them a supply chain risk. It's a company that's about to IPO that's in, you know, the sort of cutting edge of the AI race, which the Trump administration says America must win. Is that a designation that you feel should still be in place?
Oh, yeah. I mean, if their philosophy written in contract and in their constitution says that the model can decide to do what it wants to do and override human judgment, it violates my first principle that I talked about, which we have, is that human judgment has to prevail. So you can't work with a company in the sensitive ways that we're trying to work if they say the model could just overrule the human operator. That's not acceptable. So that makes it a supply chain risk for any other product that comes into the War Department that's built using their software.
But there was a period. So Palantir was allowed to continue to use Claude in Maven Smart System for a period of time. I think it ended in September. Like, was it harder to quit that, to quit Anthropic, than you thought?
No, we got it all done within six months. And by the way, a circuit court affirmed this decision. So I think that was good news to show that we were on target there. What's hard about putting super intelligence in the War Department is that I have to put on a classified data center, and that actually takes some work. So for all these companies, you know, they haven't done it, but we have OpenAI running now, we have Grok running now, we'll have some of these open source models now.
Yeah, I was going to say, which other providers are actively doing classified work now, and what kind of, you know, what...
