Eric Schmidt's argument for why machines end up running combat is not about ambition. It is arithmetic: "Humans can't run a swarm of a thousand drones, but a computer can." "So what I think will happen is computers will develop war plans within a swarm. So I'm trying to be as technically correct as possible. It will use reinforcement learning to learn how to best attack the enemy, the target, or what have you, and that those algorithms will be super secret." Schmidt separately describes a stabilizing standoff as his hope, and only a hope: if both sides have systems this powerful, perhaps neither will attack because neither can know how the other system will respond. Schmidt is the former CEO of Google and former chair of the US National Security Commission on Artificial Intelligence. He is also, as his interviewer notes in the same program, a supporter of Ukraine who has helped with drone technology, so he has a direct stake in this subject. This was published on September 8, 2026. The prediction and the hope should not be collapsed. Schmidt predicts computer-generated swarm tactics learned through reinforcement learning. He hopes their power and secrecy produce restraint. The first is his forecast about how the systems will work; the second is his proposed outcome. The supported mechanism is that a thousand-drone swarm exceeds direct human operation, so computers would generate and execute tactics inside the swarm. In the next exchange, Richard Engel quotes Schmidt's prior "humans on the loop" framing: supervising, auditing, and intervening, but not authorizing every individual shot. Schmidt then confirms that a human would still be watching. The question is therefore not the total removal of people, but whether that higher-level oversight remains meaningful when decisions occur at machine speed and the learned tactics are secret. That is the control problem in this clip. Higher-level supervision is not the same as approving every action. If military AI is expected to make tactical decisions at swarm speed, the ability to monitor and intervene has to be tested against that speed before deployment, not assumed from the presence of a human observer.