chrisl

ML is not a subset of methods that are explainable, but there are a subset of ML methods that are explainable. At least in the sense that you can tie them back to specific parameters and physical models. Not all ML methods are, but you can do ML within those constraints. And you have to when you're trying to convince scientists that they should use ML to filter large data sets where they will likely never see the data that got filtered out. You don't need the ML to do much besides identify the mines. Flying is deterministic and doesn't need AI or ML (and isn't really a good application for them). Even coordination can likely be done just fine with relatively simple deterministic methods without being demanding compute-wise. As far as I've seen from most "AI" methods is that they're reducible to enormous multi-parameter fits, with similar pathology. It's not a "mystery" but AI methods have bad habits of looking good on some data and terrible on other data that are only very slightly different (extrapolation outside the fitted space, rather than interpolation within it). And because it doesn't really give you the details of how it came to that result, and it's hidden in the parameters of a giant fit, it's not particularly explainable.

Something crossed my mind after writing the reply about the ammo dumps possibly having explody stuff moving in faster than it moves out. That's probably telling us something else, too. Explody stuff (and big explody stuff, not just boxes of bullets) is accumulating in the near-ish the theater ammo dumps and it's not making it to the lines. That seems to suggest that either Russia has a logistical problem if it's not getting that stuff to where it should get used, or it doesn't have enough delivery systems to get the big explody stuff to where it will do something besides be visible from the next planet over when it blows up. Are they accumulating a lot of artillery shells that aren't getting fired for lack of tubes? Rockets without launchers? Kilotons of explosive stuff in reserve and they have to send in meat assaults seems like a major disconnect in operations.

Por que no los dos?

I work with a Serbian woman who speaks perfect english. But I began to suspect she learned it from Scooby Doo when she correctly used "bamboozle".

Probably the same way that the LAPD bomb squad blew up a neighborhood while disposing of some homemade fireworks. Bad (or lazy) arithmetic and failure to follow procedure overloaded the bunkers. LAPD didn't bother using a scale when loading up their bomb disposal truck and grossly underestimated the amount of explosive they put inside. Then they detonated it and it blew up the truck and neighborhood. The Russian military hasn't shown that it's all that great at following the kind of procedures that one does to maintain things at storage bases, so it's entirely likely that they've just been shipping tons and tons of materiel to the staging dumps. The people at the dumps just stick it wherever. Even if it started within spec, if stuff keeps shipping in faster than it ships out, then they have to put it somewhere. Overfill the bunkers then store it between them. Then boom.

Fiber is very different to deal with than wire. And bare fiber is very different than jacketed fiber. I got called in to fix someone's fiber problem where they were breaking fibers left and right and having spools explode on the shelf. It was all bare fiber. I substituted fiber with a decent jacket and after than broke only a single fiber in I think 3 years and that was basically by accidentally doing the equivalent of smashing it into open scissors with a hammer.

I'm distinguishing them by ML being something where I can use a lightweight algorithm for the tracking, even if it takes a lot of training. I also live in a world where ML is expected to be explainable. Deterministic(-ish) once trained. I already do similar tracking with ML. AI is a bucket of mystery that sucks energy and looks impressive on some problems and falls hard on its face on similar problems that aren't easily distinguished from the ones it can do.

When I look up the part number of the second link it gives me "Flexible steel tube armoured optical fiber cables". That's not gonna be 150 grams/km. It didn't give mass specs, but looking up other armored fiber optic, it's going to be more like 1.5-1.75 kg/km for a single fiber in an 0.5 mm stainless tube (what it looks like that P/N is). And it's going to be a lot larger spool and a lot stiffer cable to feed out. So I'm skeptical that they're anywhere near 10 km with fiber tethered drones in a real environment.

I checked the date on the article and it's early Aug, but I swear I saw that picture of the fishing reel full of fiber 1000 pages ago. Bare fiber is amazingly strong until it's not. Try to break it by bending with your fingers in a really clean environment? Super resilient. Scratch it and then bend it? sproing, now you have two pieces. Get it wet or put it in a high humidity environment and scratch it across some rocks? Now you have a bunch of pieces. Leave a spool of it on the shelf in a humid environment too long? Sproing, you have a fiber wig.

I was reading further along through the thread to see if this came up before I posted. Yes, it's easier to detect them if they move. You don't even need continuous coverage, so much as repeated coverage and the ability to coregister images. If you can do continuous coverage it's much easier. Optical detection of something that's low contrast and at the resolution limit is iffy (though it does seem that in some conditions and some wavelengths, mines have high contrast). Detecting moving things is *much* easier - you're looking for changes. You can detect things that are both low contrast and below the Rayleigh resolution limit if they're moving. Re 1: It's almost certainly easier to make sure enough of your drones survive to clear a lane than it is to make sure your one big heavy metal thing survives. More advanced drones should be able to do things like come in low, or through coverage, or even come in under cover of smoke. Re 2: It's not really even AI. ML can do it, especially if it's moving, but probably also if it's not. ML might be susceptible to decoys. Why would you use decoys, when more mines probably cost about the same? because packing a lot of mines close together risks having one blow up its neighbors. But it you have a lot of things that look like mines to the swarm, but only some are real mines, it makes it more expensive or slower or both for the breacher to clear the lane.

If I can do all that (which is about as far into the future as Max Headroom), why would I bother with an expensive breaching vehicle that I can only afford a few of? I could spend the same money on a several pallets of FlexiDronesTM and install the mine clearing package on them back in my mom's basement, then transmit the locations of the target mines to the lot of them just before delivery to the breach. Press "Go" and have the overwatch drones keep track of who blows up their mines and who doesn't, then send out a second wave to clean up. It's not a lot different from a drone light show, except the drones are all carrying shaped charge detonator packages to blow up their target mines. Way less subject to the single point failure mode of a small number of big expensive vehicles.

Norm Augustine published a bunch of plots showing the extrapolation of things like this in "Augustine's Laws". One of them is that by 2050 or so, the USAF will be one aircraft and it will cost the entire US GDP. Tanks are a bit cheaper than aircraft, so the US army won't be reduced to a single tank until maybe 2080.

I didn't see any ducks imploding last night.

MS? They'll never be able to do it. They can't even show me the right emails in a search, and have a habit of trying to push obsolete documents into meeting invites as if they're current attachments. MS is stunningly bad at picking relevant things, terrible at search and indexing, and will cause no end of complex system failures at other companies. Anybody counting on MS for targeting intel should expect something even worse than when Milo Minderbinder contracted to have the US squadron on Pianosa bomb Pianosa because it's more economical that way. And when the syndicate profits, everybody benefits. Because everybody has a share. Amazon will be the ones to do it. They look at my browsing and purchasing patterns and dispatch a truck with things I might want in the direction of my house as soon as I pull up the amazon page and before I even type the search term. When I press "purchase" they just pull the correct item out of their overbroad predicted list, slap a label on it, and drop it on my porch.

"weren't told" and "didn't know" are two entirely different things, separated by plausible deniability.