> What makes you think that AI can't incorporate those into decision making?

The fact that state-of-the-art AI already fails at much simpler decisions.

In the case of the Black Sea incident, the potential consequences include global thermonuclear war.

The same AI that gets confused when you stick a traffic cone on the hood? Yeah, I don't want that algorithm deciding who to bomb.

Self-driving cars are a much harder problem than anything airborne.

Maybe we can use autonomous drones to shoot the traffic cones off the self driving cars then.

Wrong. Flying from point to point is easy. Following complex ROEs, using combined arms tactics, dealing with system failures, identifying valid targets, and employing weapons are all much harder problems than self-driving cars.

It's always hilarious to see the confidently incorrect comments by a bunch of ignorant software developers. The Dunning–Kruger effect is on full display here.

LOL, OK. You've listed a lot of problems that have been largely solved already, and are trying to convince us that they are harder than a problem that has eluded the brightest people in the tech industry, armed with computational tools that the aerospace community never dreamed of and backed by more-or-less infinite capital.

Nothing is harder than self-driving cars. Nothing. We'll colonize Mars before we have a solid solution to that problem. Why? Self-driving cars have to coexist with human drivers and human infrastructure. Nobody in aerospace has that problem. If they did, they'd run screaming for the hills.

You've been watching too many movies and are just making things up. Those problems haven't been solved in tactical aviation.

No, it's just that every time I drive somewhere, I try to maintain a low-priority thread in my head to work on the problem, "How would I write code to do what I just did?" Frequently the answer is, "I have no idea, and wow, I'm glad it's not my job."

That simply doesn't happen when I fly my quads. "How would I write code to dodge an attacking drone? How would I modify my drone to drop a grenade or a Molotov cocktail, or otherwise cause a large amount of grief to people below? How would I build a SLAM model that allows the drone to do this without intervention from the ground?" None of these engineering problems bug me the way driving a car would. They are all addressable with multiple degrees of freedom, both literally and figuratively.

Meanwhile, on the road:

"Hmm, the light at this intersection is out. There's a cop with an angry look on his face, flapping his arms at me like a dying chicken. What does he want me to do, exactly?"

"Huh, here I am in Seattle, and it looks like they have chosen to mark the stripes on the road with some sort of paint whose complex impedance at optical frequencies is identical to that of rainwater. I'm sure glad I'm driving, and not my lane-keep assistant, which I had to turn off because it tried to steer me into the median the last time it snowed."

"Whoa, where'd that ambulance come from. The law says I have to move right, but the only way I can get out of his way is to move left, and in any case, that's what the car ahead of me is doing. What to do, what to do."

In most of the airborne scenarios you mention, doing nothing is a perfectly viable, fail-safe solution when confronted with a situation the hardware or software can't handle. If we approach driving that way, a few miscreants can brick an entire city, intentionally or otherwise.

I'm not surprised Karpathy tapped out, let's put it that way.