Palmer Luckey argues that the main reason to put decision-making software inside a weapon has little to do with speed. The reason, he says, is that the radio link back to headquarters may not survive a war. Luckey founded Oculus VR and the defense technology company Anduril Industries. He made the case on stage at Moonshots Live 2026, in a conversation recorded on September 25 and published as an episode of Moonshots with Peter Diamandis on September 28.
The question: headquarters or the field?
Dave Blundin, founder and general partner of Link Ventures, framed the issue in computing terms. When drones replace pilots, he asked, who releases the weapons, and where is that decision made? Sending sensor data back to headquarters gives access to far more computing power, but it adds delay. A chip inside the aircraft or robot is fast but limited. Would the future bring "self-contained decision-making weapons wandering around enemy territories," he asked, or control from headquarters?
Luckey said the real concern "isn't so much the latency or the speed of computation." Relying on a central site to do the computing means keeping a fast, reliable two-way connection with potentially thousands or tens of thousands of machines in the field. If an enemy jams those communications (floods the radio channels with interference), intercepts them or hacks them, "you're done," he said. Worse, the enemy can ignore the machines in the field and strike the place that controls them all.
He illustrated the point with a joke about "a documentary that George Lucas was working on." In Star Wars: The Phantom Menace, the Trade Federation's central command was destroyed. That, Luckey said, was how the battle of Naboo ended: the battle droids lost. "Oh yeah, they all fell over," Blundin replied. The official Star Wars reference describes the Droid Control Ship as the vessel that transmits signals to the battle-droid armies. In the fictional example, many machines depend on one signaling platform.
Luckey gave other reasons for moving the thinking forward. Underwater, you often cannot communicate at all. Even where you can, radio transmissions can reveal your position. "For better or for worse, you're going to need to do this," he said.
Anduril's developer documentation for its Lattice software platform describes the same design choice. It lays out a local-first approach. Early Sentry towers ran computer vision and combined their sensor data on the device itself, and applications are built to work with limited bandwidth and on networks that are sometimes unavailable.
"That's been the case for decades"
Luckey said critics who warn that a "Pandora's box is being opened," with weapons making their own kill decisions, overlook how long such weapons have existed. He said the Aegis system on destroyers has a fully autonomous mode: flip a button and it will shoot down anything that comes within the range it has been programmed with, without human confirmation. He said the same is true of close-in weapon systems and of C-RAM (counter-rocket, artillery and mortar) systems. These are used against incoming missiles, mortar and artillery fire, and speedboats.
The best-known close-in system, Phalanx, fits part of that description. Its maker, Raytheon, describes it as a self-contained, radar-guided gun that automatically detects, evaluates, tracks and engages threats. On ships it is a last layer of defense, and a land-based version intercepts rockets, artillery rounds and mortars. Raytheon's description also qualifies Luckey's claim. For some targets, including helicopters and fast boats, the Block 1B version adds infrared sensors and operator stations so that people can identify a target visually before the gun engages. The system is also old: a Pentagon test report dates Phalanx testing to the early 1980s. The same report found a later version effective against certain anti-ship cruise missiles, but called its effectiveness against others only potential.
Luckey then went back to the Vietnam War. Missiles fired from aircraft flew beyond the horizon, searched for radio emitters believed to belong to a surface-to-air missile site, and struck them. The human role came before launch. A person decided to send the missile into a particular area to look for a particular signal, knowing the probability that it would get it right. "That decision is going to remain with people," Luckey said.
In Luckey's account, human authorization means choosing where a weapon operates and what it looks for. It does not mean approving each shot.
Traps, mines and golems
Luckey took the history back much further, in what he said was a rant for another talk. He defined autonomous weapons broadly, as weapons that carry out what they were programmed to do independently of the operator or the engineer. By that definition, he said, they have existed for thousands of years. One example was a pitfall trap that lets people walk over it but catches a commander on horseback, because the commander and his horse are heavier. Another was a poison trap meant to kill people with certain characteristics but spare others, or spare animals. He also listed many kinds of landmines and sea mines, and even torpedoes.
Luckey acknowledged that the claim "depends on your exact definition." He extended it to myth, where people create golems that fight on their behalf or defend a city. "We've had autonomous weapons for a very, very long time," he said. "It's just a new set of technology that's making them better." He was not equating a pit or a golem with modern AI. His definition groups them because each carries out what someone set it up to do, with no operator stepping in at the moment it acts.
Not a chatbot in a missile
Blundin followed with a technical question about the foundation models, the large general-purpose AI models behind chatbots, that would run in these remote systems. He said the open-source models available right now were all Chinese, which would not go into remote weapons, and asked how Anduril would get AI into its devices.
Luckey said the company had been doing this since 2017 and was "a decade beyond that question." In his view, the foundation models people are building are the wrong kind for the job. They are "largely non-deterministic," meaning the same input can produce different outputs, and "non-auditable," meaning their decisions are hard to inspect afterward. He said that what Anduril builds has very few consumer applications, and consumer models have little use in his systems.
He did not dismiss language models in defense. There are "LLM use cases for defense that make a lot of sense," he said, "but it's not guiding my missile" or classifying radar feeds. Anduril's developer overview describes that kind of division of labor. Lattice's data services include text tasks such as translating language and summarizing a tracking picture, alongside services that correlate sensor tracks. Anduril chief executive Brian Schimpf has told Congress that Lattice combines observations from many systems into an operating picture for human decisions.
For weapons, Luckey said, the input is not text. "We can't build models that ingest text and then spit out text," he said. His systems need to process raw radar feeds and sonar data from passive listening stations on submarines, work he called "pretty highly specialized." He also took some credit at the expense of AI's newer enthusiasts, most of whom he described as recent converts. The initials of Anduril Industries are AI, he pointed out, and "that was not a mistake."
The risk he worries about
Earlier in the conversation, Peter Diamandis, the founder of XPRIZE, had asked Luckey for his "P-Doom," his estimate of the chance that AI causes catastrophe. Luckey's answer matched his view of weapons as tools that carry out human choices. "I'm not worried about AI doing the killing," he said. "I'm very, very worried about evil people using moderately competent AI to do evil things."