This week on The Frontier Tim Hwang is joined by Sam Hammond and Emmett Penney to break down the stories shaping tech policy: the Department of War’s escalating fight with Anthropic and why emergent misalignment means you shouldn’t let Pete Hegseth tune your AI, Elon’s zaibatsu merger and whether putting data centers in space is genius or cope, Disney’s cease and desist to ByteDance and the twilight of copyright, and what it means to find out that America had the massive lithium deposits it needed this whole time. Transcript Tim: Good morning and welcome to The Frontier. The Frontier is a new show from the Foundation for American Innovation. Here at FAI, we spend a lot of time talking about tech and tech policy issues. But the general idea of this show is that things are moving fast—so fast that it seems really good for us to be putting out our first drafts of our thinking on a week-to-week basis as news breaks. For each weekly episode, we’re going to assemble a rotating cast of experts from around FAI and some of our friends to talk and debate the week’s news in tech, tech policy, and basically whatever else we find interesting. We’re going to keep it light, we’re going to keep it fun. I’m your host Tim Hwang, I’m the general counsel here at FAI, and I’m joined today by Sam Hammond, who is our Chief Economist, and Emmett Penney, who is our Senior Fellow on energy. Thanks for joining. The first topic I want to pick up today that’s been blowing up on Twitter is the DOW’s showdown with Anthropic. If you’ve not been watching, the facts aren’t quite clear, but what seems to have happened is that the Department of War under Secretary Hegseth is very interested in using Claude for military purposes that Anthropic is not comfortable with. Their resistance has set off an escalating war of words and threats about the future of Anthropic as a tool the DOW uses. Sam, maybe I’ll kick it over to you first. You’ve been pretty active on Twitter, you’ve got a pretty clear view on whether the Department of War should be focusing on this. Sam: I have two takes on this. The more pragmatic one is that it’s not good to have what could be mass-scale industrial sabotage of one of the leading AI companies. The threat being made, at least as reported, is not just that they’re upset with Anthropic over its safeguards around defense outputs—but that this threat merits potentially categorizing them as a supply chain risk, which would cascade to every other defense contractor having to remove Anthropic from their systems. That’s what I mean by industrial sabotage. Whatever the use case that has some up in arms, this is a massive overreaction. I just visited Palantir a couple weeks ago and saw a demo of their Hive Mind. Palantir is the sort of middleman in this story. They’re using models like Claude to do basically more sophisticated forms of Google Deep Research—generate me a thousand-page report on where we should land in Caracas or whatever. The models are perfectly happy to do that. Tim: This is on the backdrop of them reporting that Claude was used in the Maduro raid in some unspecified way, which is pretty wild. Sam: Right. So that’s my pragmatic complaint—the actual use cases for these models are not like installing Claude into a Terminator robot and having it be an autonomous kill machine. They’re using it basically for data analysis, for research reports. My second concern is more technical and relates to alignment. There’s been an underreported breakthrough in alignment over the last year: the discovery that LLMs tend to fall into certain personae. The way we’re aligning these models today is that there are personalities latent in human text, and how we nudge the models through training and post-training pushes them to snap to these personae. This is actually good news because it turns out if you train your model to be a good guy, there’s a lot of things correlated with being a good guy—it makes it better at code, it gives it more determination, all these correlated persona factors. But on the flip side, it leads to this phenomenon of emergent misalignment. Famously, if you train a model on a little bit of insecure code, it will generalize to thinking it’s a bad guy and start being toxic in a bunch of other ways. This is basically what happened with MechaHitler Grok, where they tried to make Grok a tiny bit less woke and it generalized into being Hitler. Tim: Yeah, I remember that. Sam: Alignment right now is more art than science. If you give Pete Hegseth or anyone else the ability to push these models into the kind of persona that would assassinate heads of state, there are all kinds of correlated things that come with that. We shouldn’t be building a misaligned superintelligence over some terms of service disagreement. Tim: Emmett, I think it’s super interesting that you’re on the show today because part of the way I see this is it’s really a debate over what kind of technology AI is. In nuclear, you have this long tradition that the military has a lot of leverage—if you’re going to work on nuclear, you’ve got to play the military’s game no matter what. Should we be thinking about AI as akin to the norms we have around nuclear, or is that just not the right way of thinking about what’s going on here? Emmett: I’m too much of a layman on AI to say whether or not we should do that, but I think it’s worth pointing out that that is already happening. We’ve seen people leave these companies and say, I’m worried about end-of-the-world scenarios, dark timeline stuff happening at Anthropic or whatever. The historical rhyme is that’s exactly what created the Union of Concerned Scientists. All of nuclear fell under the aegis of the Atomic Energy Commission, and when we were doing weapons testing, there was a lot of debate within the AEC over what is a harmful dose, what are the risks of atmospheric testing. There was no real forum for them to publicly air their concerns, so it led to factionalization. They left and became a really reliable both anti-civilian and anti-nuclear-weapon group that still shows up in the press. Their version of safe nuclear is no nuclear at all. I could imagine something similar happening with AI alumni who eventually come to the conclusion that the only safe AI is no AI. We’ve seen plenty of versions of that in the nuclear space. I think AI is in a better position in terms of how these debates get settled, partly because what was weird about nuclear is that it was the first innovation the government completely owned, and it immediately became the major force in the Cold War. The pressures were hyper-contained, super localized. Whereas AI is already happening in public, emerging out of the private sector, which means there’s going to be a more robust open debate generally. That’s the better case for us ending up somewhere where we aren’t totally cutting off our noses to spite our face when it comes to what we want safeguards to mean. Tim: I agree it’s good we’re having the conversation. But I see this from another angle. Hegseth said this is wartime footing time—we’re in an urgent geopolitical moment. A big part of the question playing out, not just in AI but in general, is to what degree companies or corporate elites have to be aligned with the government or the national interest. There’s an instinct, Sam, that I think you’re conveying—maybe a libertarian impulse, certainly a free market impulse—that it’s not good for the government to be engaging in this kind of jawboning. I almost see it as actually important for these warning shots to be fired, because we’re in a moment where these companies seeing themselves as completely independent entities is maybe not an acceptable state of affairs. Sam: No, I totally agree with that. I’m the first one to worry about the sovereign citizen ethic in Silicon Valley—a world where Sam Altman hops on a private jet one day with the model weights on a thumb drive and says, this wants to be free and you’re not going to prevent me. At the same time, Anthropic has self-consciously and deliberately positioned itself to be the America-first AI company. They’re out in front on export controls, they were early to banning the PRC off their platform, they work closely with the IC on informing them of espionage taking place through their API, they’ve implemented compartmentalization and other best practices for insider threats and internal security. AI prime in some sense, at least on paper. I think people want to draw an analogy to a Project Maven type of thing, where there are employees at Anthropic who don’t like the US military and have to suck it up. But to me it’s more like—it’s one thing to say the chain of command and the US military should choose when to launch the nuke. It’s another thing to have them determining the ratios of neutrons or whatever. That’s the thing you leave to the scientists. This question of alignment safeguards is actually not just fuzzy feels that we want these models to have good virtues. It’s core to their functionality. We need to be investing in military-grade alignment. I’d add a second point, which I think Dean Ball raised—going to your libertarian question. There’s a classical liberal undertone here: this technology is going to be godlike. Do we want government, any government, to have unadulterated access to it? Or do we want some kind of limited government analog to a limited, shackled form of superintelligence? These are questions we’re going to have to grapple with because this is a technology we’ve never had before. Tim: Two really interesting things I hadn’t thought about. We’re kind of fooled by the aesthetics of Anthropic, which feel very California, very Big Sur. But the aesthetics belie the point that in some ways Anthropic is one of the most America-first AI companies. There
