As someone that works in ML, this is actually what is concerning to me. Everyone is talking about AGI (artificial general intelligence) but I don't think that's something of huge concern yet. We have already entered a world where you can create drone "mines". It is cheap and easy to build a drone that can have an explosive payload, hide, and automatically seek out enemy combatants or vehicles. (Note that drones are pretty difficult to detect) The tech is a little difficult now and requires oversight if you don't want to violate international war laws, but it is definitely possible (and rapidly getting better).

> If one $50k drone can consistently take out millions of dollars of equipment, it doesn't matter how expensive or numerous that equipment was, or how good it would be at fighting some hypothesised similar adversary.

Because this isn't true anymore. It is really a $1k drone being able to take out millions of dollars of equipment with a 70+% success rate. That's a real game changer.

We don't need AGI to for ML to be dangerous. We just need people to use existing algorithms dangerously and/or recklessly.