Palmer Luckey: Autonomous Weapons Are Ancient and Why Anduril Won't Build Humanoids | Moonshots Live | EP #295

Palmer Luckey’s argument is that autonomous weapons are not a new category: under a broad definition, they include weapons that carry out programmed intent without a person making the final engagement decision. Traps, mines and torpedoes fit that definition, as do modern systems such as Aegis and CWIS/CRAM. He also points to Vietnam-era missiles that could engage without confirmation at the moment of attack. In that example, people still chose the operating area and the signal the missile would treat as a target, accepting the risk that it would identify the target correctly.

For Luckey, autonomy is often a practical requirement in the field, not simply a way to speed up decisions. Underwater systems may lose communications, and a vehicle that transmits radio signals can reveal its position. Remote control also creates vulnerabilities: links can be jammed, intercepted or hacked, while a central control site can itself be attacked. The trade-off is therefore not just between human judgment and machine speed; relying on communications can expose both the system and the people controlling it.

Luckey says Anduril has worked on autonomy since 2017, with defense AI designed for specialized processing of raw radar and sonar data. He rejects the assumption that consumer text-in/text-out foundation models are generally suitable for these missions, describing them as nondeterministic and unauditable for the required tasks. Defense systems, in his view, need to interpret sensor data in ways that can be assessed and trusted, rather than simply produce plausible outputs.

He also doubts that defense platforms will converge on one general-purpose robot. Different missions call for different equipment, and even a small performance advantage from specialization can matter when facing an adversary. Anduril has chosen not to build humanoids, small quadcopter attack drones or space-launch vehicles. Luckey expects defense platforms to remain specialized, including deep-diving underwater robots and separate surface and aerial vehicles, rather than treating one adaptable design as the answer to every operating environment.

Humanoids may make more sense in manufacturing, hospitality and medicine, where economies of scale can outweigh the performance advantages of a task-specific machine. Luckey also sees them as a possible interface for AI controlling remote systems or legacy platforms built around human controls. That potential does not change Anduril’s decision to focus its defense work on specialized platforms; it reflects a different use case, where compatibility with human-designed environments may be valuable.

The company’s autonomous firefighting tank illustrates the demands of a specialized system. Luckey says Anduril began building it in 2017 and finished it in 2019; the project won a $2.2 million wildfire X Prize, including $1 million from Lockheed. The system has to detect and classify fires, distinguish them from false alarms such as campfires or road flares, choose an appropriate nearby asset and act before a fire grows. Timing affects what response is useful: Luckey says a 10-minute response can require entirely different tools from one that begins an hour later.