“I think A.I. is a Trojan horse that everybody knows the Greeks are inside. I’ve never seen a technology advancing so rapidly, so completely rejected by the public. Technology is always going to give us great gifts, but it has to be viewed with skepticism; the motives of the people giving it to us also have to be viewed with skepticism. That’s when we’ll get the best out of a new technology, rather than blind faith that everything’s going to be great.” —Christopher Nolan, a few days before the $124.5 million opening weekend for The Odyssey, just in case you needed another way to interpret the film’s themes.

And now for the main event…

Shortly after Russia invaded Ukraine in 2022, Putin’s forces began assembling a series of complex defensive fortifications known as the Surovikin Line—minefields, trenches, ditches, and concrete “dragon’s teeth” covering dozens of square miles of Ukrainian territory. For a time, the two countries were locked in a horrific war of attrition that seemed like it would inevitably end in victory for Moscow. Corey Wilkens, a sergeant major with the U.S. Army combat engineers, recalled the intense trepidation he felt studying the Surovikin Line from his post at Fort Bragg. “Not only would we have a really, really hard time being able to breach that in general, but with our current methodologies, we would have mass casualties of our engineer units if we had to go do that,” he told me.

By January of 2024, however, Wilkens’ analysis had turned into the Sandhills Project, a well-funded program led by Wilkens and focused on the development of unmanned systems designed to remove soldiers from the most dangerous parts of combat operations. In this case, that meant breaching defensive fortifications. In March of 2025, the U.S. Army awarded a $114 million contract to Forterra, a military contractor focused on autonomous vehicles. By October, Forterra had supplied the Ukrainian military with 105 “Lancer” autonomous ground vehicles, built on top of the Polaris Ranger 1500 A.T.V. platform.

The Lancers—nonlethal self-driving vehicles capable of going off-road, gathering data, and trucking matériel across the front—have since been used across more than 1,100 missions, including 52 casualty-evacuation missions.  All this as Ukraine has become the vanguard for A.I.-enabled defense capabilities supplied by allies—maritime drones, self-driving ground vehicles, and a host of A.I.-enabled targeting systems for both air defense and ultraprecise, low-cost drone strikes. In a statement earlier this month, a Ukrainian commander described the Lancer as “the number one choice for critical logistics missions,” adding: “In fact, we urgently need more Lancers to be shipped over immediately.”

Getting these systems ready for use was no easy task. “Ground autonomy is hard,” Wilkens told me, and even harder in an active combat zone. As Forterra’s chief growth officer, Scott Sanders, told me, “I want the vehicle to do exactly what I tell it to do, because it’s going to go play in a minefield. And it’s really hard to use agentic A.I. for stuff like that because there’s nothing to train it on—no one’s ever done that before.” The solution, he explained, is an autonomy software that uses a mixture of classical and modern robotics techniques to grant the vehicles’ operators a kind of semiautonomous controllability. “In certain parts of certain missions, you don’t want it to try to figure out what it should do,” Sanders explained.

The resultant vehicles operate with a higher acceptable margin of error than, say, Waymos. That makes sense: There’s a greater risk tolerance in combat than on the streets of Austin. Even a system that works half the time, Wilkens explained, keeps soldiers out of the breach. “You’re dealing with a more unpredictable environment,” as Josh Wallin, a defense fellow at the Center for a New American Security, bluntly put it. He described the semiautonomous nature of these systems as a “pretty natural next step, and it also seems like one where you’re less likely to end up with casualties.”

And while a driving force behind the project was to better equip the Ukrainian military, the organizations behind it are clear-eyed about the benefits of having combat-tested technology ready for U.S. use, as well. “That’s how you don’t show up to the next war with the wrong army,” Sanders said.

The current version of the Lancer, Wilkens said, is “incredibly different, from a software and capability perspective,” than the original. Still, he told me, the systems are not yet “using a lot of A.I.” That’s an experiment reserved for Sandhills 2.0. “Our next step, the framework that we just drew out, is to incorporate A.I. and agentic A.I. into our processes,” he said.

Wilkens described an agent, for instance, whose sole purpose would be mission planning for the autonomous ground vehicles. “As long as we’re staying very narrow, that’s all possible today,” he said. “It is pretty easy, and we still have final control and approval of the mission. It would rapidly increase our ability to be able to push missions through, and reduce cognitive load for us to be able to go do that.” Eventually, the goal is to deploy multi-agent systems, with an agentic “orchestrator” that has access to the “overall mission plan,” he said.

That timeline, of course, raises plenty of questions, ranging from military theories around the potential for robotic deescalation (or greater escalation), to the consequences of unreliable algorithms, to possible human rights violations. As safety engineers have been telling me for years, there’s no evidence that L.L.M.s are fit for military applications. And as Peter Asaro, an expert in the ethics of military robotics, told me: “All of this technology could eventually be weaponized in various ways, and so that is also very concerning.” The concern is certainly not unfounded, given Defense Secretary Pete Hegseth’s directive that the U.S. military must become an “A.I.-first warfighting force”—not to mention his clashes with Anthropic over guardrails for autonomous weapons systems.

According to Wallin, though, we’re not quite there yet. “I know the department talks a big game about this, but I don’t really see any sort of near-term development where we’re going to be deploying an L.L.M., or something with that degree of generality, in an actual use-of-force context,” he told me. “I would actually say in the case of the Department of Defense, I think there’s been a lot of reticence in developing and deploying these systems in some cases, because no one wants to be the first one to fail. No one wants to be the first one to have an autonomous robot go off and kill innocent civilians.”

Instead, the focus remains on integrating A.I.-first strategies in intelligence analysis. “Over time, trust will be developed and, I would say, calibrated, so that hopefully you have a good understanding of what tools are and aren’t capable of, so they actually get to leverage them to the fullest degree but not overtrust them,” Wallin said. “I think there is a healthy level of skepticism among folks right now about the potential for failures. I would actually argue in some cases there may be more under-trust than over-trust.”

That’s all for today. I’ll see you on Thursday.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006