Notes

The weaponization of AI is progressing rapidly due to the prolonged Ukraine-Russia war. Defense research would continue even if there was no war, but the biggest reason for faster progress during war is quick feedback from actual situations.

Again, I feel that two things are important in technology development: 1) data from actual customers, and 2) a fast feedback cycle.

Human-in-the-loop vs. Human-on-the-loop: In the former, a human is included in the decision-making loop, and in the latter, the decision-making loop runs completely automatically, but a human monitors it and intervenes when necessary. The latter is also called supervised autonomous.

The positive and negative aspects of war automation.

Positive: Reduces collateral damage (e.g. civilian casualties) and causes less mental stress for soldiers because humans do not kill directly.
Negative: If human intervention in war is reduced and killings are carried out through “cold calculation,” the likelihood of casualties and violence becoming normalized increases.

Project Maven: A U.S. Department of Defense project that started in 2017. An attempt to apply AI and ML to military technology in collaboration with Google. Halfway through, Google decided not to continue the project. After searching further, I found that the project is still in progress even after Google’s withdrawal, and (as expected) Peter Thiel’s Palantir is actively participating.

Military AI technology is developing rapidly. Recently, a dogfight simulation was conducted between AI-controlled F16 fighter jets. (The simulation referred to here is not a simulation in a computer, but a simulation in the sense that there are real airplanes flying, but it is not a real war.)

Among the military concepts, there is something called OODA loop (observe the situation, orient by analyzing information, decide, and act). For example, if observation, analysis, and execution are automated with AI, humans will be able to spend more time making decisions.

Generalization: Think about what to automate and why, not whether to automate or not.

Just like MAD (Mutually Assured Destruction; Mutually Assured Destruction. “If you use nuclear weapons, I will use them, and then we all die”) in a nuclear war, the concept of MALoC (Mutually Assured Loss of Control; Mutually Assured Loss of Control. “If automation continues to respond to each other more quickly, eventually humans on both sides lose control”) exists in military AI.

The concept of singularity on the battlefield can be considered. The point at which AI’s OOAD speed exceeds human understanding speed in a human-”on”-the-loop situation. In this case, humans effectively lose control over war.

Lesson learned: A field-specific concept of singularity would be more beneficial than a broad concept of singularity.

If the current situation continues, it will soon become similar to HFT (high-frequency trading) in stock trading. In HFT trading, even 1/1000th of a second difference is important, so human-in-the-loop is impossible in principle. The big problem is that there are control devices (e.g. circuit breakers) in the stock market, but there are no such things in war.