Global military buildup

[Syt]

As long as e.g. facial recognition is having issues with more heavily pigmented faces, I'm loathe to trust an automated recognition of enemy fighting vehicles and people.

For vehicles I'd assume that thermal cameras are more practical, as a running engine is going to be much hotter than anything in the background (unless you're in an extremely hot environment). Now, I doubt that a camera can tell apart a, say, supply truck, from a tank.

Supply truck and tank is easier, I suppose, than a truck with infantry in the back vs. a school bus, or a truck with crates of beer.

[Threviel]

If air supremacy is achieved and there's a total war going on you might have drones circling roads and rail tracks automatically shooting everything that moves a few km behind enemy lines.

Very useful for interdiction in a total war scenario. Not so useful if there are friendly civilians behind enemy lines or a not total war scenario.

[Threviel]

Also very useful if you are an evil dictatorship conducting a war.

[crazy canuck]

I don't think Russia or this war tells us anything about the future of Liberal democracy. After the collapse of communism Russia was never on track to be a liberal democratic state. The error the liberal democratic West made was thinking that liberal democracy was the only alternative as we watched Russia become a kleptocracy.

The future of liberal democracy is not going to be decided in this war, or Russia or China. The future of the Liberal Democracy Will be decided by whether the liberal democratic west continues to be governed by liberal democratic institutions.

I don't have a great deal of confidence that is going to continue but I have hope.

[DGuller]

Is there any technical limitation to making drones fully autonomous? I mean there is obviously ethical considerations. But technology should allow auto-piloting the drone, acquisition of targets and launching of weapons. The hardest bit is probably sensors, but identifying military vehicles seems straightforward enough.

I'd imagine all the same limitations behind self driving cars multiplied by 1000.
AI is not great at messy unpredictable situations.

I think it would be the opposite.  Self-driving cars are a much more difficult problem to solve than self-flying planes.  I think planes have been able to fly and land themselves automatically for at least 50 years.

[Grey Fox]

You can teach a visual AI anything you want. Open source solutions are already extremely accurate a differentiating between different vehicles. You just need time & computer power.

Imo, decision making is the technological hurdle right now.

[Berkut]

Is there any technical limitation to making drones fully autonomous? I mean there is obviously ethical considerations. But technology should allow auto-piloting the drone, acquisition of targets and launching of weapons. The hardest bit is probably sensors, but identifying military vehicles seems straightforward enough.

I'd imagine all the same limitations behind self driving cars multiplied by 1000.
AI is not great at messy unpredictable situations.

Actually AI is really damn good at messy unpredictable situations. 

Hell, self driving cars are not even that unpredictable. There are very easily defined limits on what they can do all based on simple physics.

[DGuller]

Another reason why I think it's not right to compare the difficulty of getting self-driving cars on the road to drone warfare is that for self-driving cars, the difficult part is going from 99% reliable to 99.999% that the public demands.  I think that in war, the tolerance for failure is quite a bit higher, for obvious pragmatic reasons, as long as on balance you're still doing more damage to the enemy than to your own side.

[Josquius]

Actually AI is really damn good at messy unpredictable situations.

Not really. It's an area where development of AI is having big problems, going beyond direct teaching and in using past experiences to decide on something completely new. This is in the path of true AI.

Is there any technical limitation to making drones fully autonomous? I mean there is obviously ethical considerations. But technology should allow auto-piloting the drone, acquisition of targets and launching of weapons. The hardest bit is probably sensors, but identifying military vehicles seems straightforward enough.

I'd imagine all the same limitations behind self driving cars multiplied by 1000.
AI is not great at messy unpredictable situations.

I think it would be the opposite.  Self-driving cars are a much more difficult problem to solve than self-flying planes.  I think planes have been able to fly and land themselves automatically for at least 50 years.

Self flying planes are easy because the sky is pretty empty and simple.
Drones by their nature deal with the ground. Which is messy.

[DGuller]

I think you need to deal with the ground to some degree when you land, and yet a passenger plane built in 1970ies could already do it.

[The Minsky Moment]

Another reason why I think it's not right to compare the difficulty of getting self-driving cars on the road to drone warfare is that for self-driving cars, the difficult part is going from 99% reliable to 99.999% that the public demands.  I think that in war, the tolerance for failure is quite a bit higher, for obvious pragmatic reasons, as long as on balance you're still doing more damage to the enemy than to your own side.

Agreed.
And in war, there is another element.  I have no personal experience in being in combat but from second-hand accounts, studies, and common sense, I understand that being exposed to lethal fire can have a negative impact on decision making. AI routines don't care if they are being shot at, except to the extent they are programmed to respond as desired.

[Jacob]

I think you need to deal with the ground to some degree when you land, and yet a passenger plane built in 1970ies could already do it.

I expect there are orders of magnitude more complexity in dealing with having to assess everything on the ground compared to executing a pre-calculated flight path in an environment where everything is mapped out to the Nth degree and where you can fall back on human operators if any variables are outside of expectations

70s aircraft landing at an airport do not have to be concerned about the movement of civilians, the application of camouflage and baffling techniques, the uncertainty about the enemy you're facing, or the evaluation of priorities (is this target of opportunity worth more than the initial target? Is the fact that the characteristics of the planned target diverges 25% from expectations sufficient to hold fire? etc) to name just a few things.

[Josquius]

I think you need to deal with the ground to some degree when you land, and yet a passenger plane built in 1970ies could already do it.

That's a very controlled situation with a pre-expected specific scenario.
All together different to interacting with the whole ground.

And really disagreed that the margin of error is less with military vehicles than civilian. Its far far more. You're directly choosing to kill people with drones. Actively choosing to blow up a convoy of refugees.

[Jacob]

Actively choosing to blow up a convoy of refugees.

... not to mention actively chosing to blow up your own people, say because you got a misread of the armbands, becaues vehicle markings got obscured etc.

[The Minsky Moment]

And really disagreed that the margin of error is less with military vehicles than civilian. Its far far more. You're directly choosing to kill people with drones. Actively choosing to blow up a convoy of refugees.

Again the relevant comparison is not against some Platonic ideal of perfect military targeting but against how human manned systems perform.  I.e what % margin of error is needed to match Blackwater contractors?