kimbosbread

Also, warm up for our equipment just to see it’s up to snuff, and a giant kick in the balls re cheap drones. That said, I don’t think the state of things is anywhere near as good as this, forgetting completely about Iran getting Russian nukes. What happens if Russia continues terror attacks? What incentive do they to not continue? Russia can continue destabilizing our political discourse pretty much without consequence. The lesson for every bad guy from this is that you might get a lot of soldiers killed and military equipment ruined, but you can seize that territory you wanted. And that might be good enough for Winnie the Flu. I bet he would happily trade half the Chinese military for Taiwan. Oh and there might be diplomatic isolation from the West, but nobody else really cares.

Regardless of whether it’s a soldier in powerered armor, or the suit without a human inside, or a robot dog, you need something to occupy territory. Occupying territory might just become very expensive.

I think we are closer to being able to rely on at least some battery power than is commonly accepted. V1 powered armor is likely just mobility, where the soldier can now jog at 10kmph with a full combat load for 1 hour without destroying their joints. Let’s say takes 1kwh. If the solider is packing along a few kg of tesla-style batteries, that starts look less sci-fi and more real. I think the major challenge is engineering a suitable actuator (ie artificial muscle) for this suit. It’s all about the economics. If 10 soliders cost as much as one tank, but each one has similar destructive capability by themselves relative to the tank, then things are ok. Yeah we have to differentiate between the various benefits of powered armor, and how much they cost. For example, improved mobility, vs more weight, vs better camouflage, etc. Ugh I’ve only finished part 1 and a week or two of 12 hour days has put a halt to more reading.

Sorry for the off-topic! I apologize as didn’t mean to be so contenious, or imply any of what I wrote is a value judgement and thus we should convert into a patriarchy with concubines. It’s just that social issues often do have major social ramifications. If the workforce size doubles without any other changes, that has many second and third order effects. How countries and people deal with this is gonna be a huge deal. Free childcare on a national level seems like a slam dunk, but how does that work if you cannot even find enough public school teachers? You also have interesting social issues like single women in their 40s choosing to become single mothers via IVF (during pandemic IIRC fertility treatment numbers were 25% lesbian couples, 25% heterosexual couples and 50% single heterosexual women, IIRC). It’s not about white people or skin color or anything else in most cases; though Russia is a counter-case where they have less “real” russians and more caucasians.

Oooooo, I get to be disagreeable. Let’s talk about women’s sufferage, and entrance into the workforce en masse. Several interesting symptoms, across many countries, from Central American Kleptocracies to Western Democracies to everybody else, pretty much: As women get educated, less children Twice as many workers, and thus many things cost twice as much that have an inelastic supply such as housing Fewer meals prepared at home cause no home-maker (otherwise known as the Elizabeth Warren Diet), with the second order effect of less healthy eating habits Women are not locked into marriages, and thus a higher divorce rate Women prefer not to date down, thus with more women going to school, there are fewer desireable eligible men, hence lower marriage rate All of these together reduce the number of children being born, which over time will cause major social upheaval. Obviously the ramifications of aging societies haven’t been fully explored yet, but it’s going to be exciting. Do we not think there is going to be social collapse in at least a few countries?

Well yes, this is one of the things that happens when the average reading level in your country is around sixth grade (because some ivory tower academic careers benefit from it), and that’s the average. Median might be scarier. Obviously it’s not entirely literacy, but numeracy is way worse.

There others who are more subtle about destroying our society! Lucy Calkins for example has done more harm than just about anyone by destroying American literacy: https://www.nytimes.com/2022/05/22/us/reading-teaching-curriculum-phonics.html On the left, I’m going to say the doors-wide-open immigration policies and the reformist-minded-approach to crime, which are both weird because in the end, they hurt the big liberal cities the most while the conservatives are just puzzled that if you wanted to live in a big city, and that’s where your voters are, why would you trash it.

We’ve got to find out what people want from UGS, how they relate to them, what sort of image they have for them. For example, do people want UGS that can be fitted nasally?

When I retire from building distributed systems and machine learning whatnots, I want to build the machine overlords!

These are two very important points worth repeating!

You have to distinguish between what exactly you are spoofing. Are you trying to hide from a human, a high resolution camera, or a crafty computer, or a combination of all 3? A human controlled fpv drone is oven NSTC (720x480px) to reduce data throughput/latency, plus there are always signal issues. So maybe you can disguise yourself enough from that. A more surveillance oriented system might be shooting 4k video, which will make harder. Or there’s thermal. An autonomous drone will be limited by onboard processing power/ml algorithms and models. Given the amount of power available on modern phones, you can get that a fused thermal + optical sensor could recognize camouflage nets and tarps, and see those as a nice juicy target. And if it’s a quadcopter, why not just sit down on the ground and watch and wait? Loitering munition doesn’t mean it has to be in the air to loiter.

Why not gasoline? Those micro-turbines and whatnot gives way better endurance. I’ll grant that batteries have other advantages however.

Yeah, you don’t need to be perfect either, as long as your false positive rate is zero. Training a model to recognize locomotives and train tracks is pretty trivial, and as discussed pages ago, this is something that should be done to crush Russian logistics. I think loitering for up to 12h and finding artillery would be the other. Just wait till it moves if you have a reasonable amount of drone density, say a few in a 10km square grid. I really think the artillery hunter and train hunter cases a few bored people could build in a few months. If I manage to deliver my big project for day job this week or next, I’ll build a model post it on here with repro for everybody’s satisfaction. Sadly 10-12h days precludes working on anything else. The economics argument is on point. Rephrase it to “Each submunition has its own guidance system and locomotion”. Not necessarily a quadcopter, but the line between drone and smart munition is rather blurry.

I was thinking of it more as an anomaly detection problem. If you could label your data using what Ukraine does (given either SAR or optical, or combined), and train the model on that, would that work on the other side as a rough cut? And then once you have located them, loitering munitions. I guess that’s the missing link: An autonomous drone that can recognize desired targets and look for them in a rough search zone for a few hours.

Yeah, the accurate labeling is a real issue. For my drone problem, the nice thing about a virtual environment is you absolutely know where the virtual drone is, so you have labelled data without extra effort and can very easily build the model to estimate distances for a single camera- once you have the environment.

Also, I’ve wanted to give everybody some insight into one of the big problems of how you train image recognition models: Good quality data! I mentioned a while ago that I wanted to set up my drone to fly through my house autonomously, exploring it depth-first-search style and building a map, basically, and driven by an ML model running on my rather beefy desktop. I realized pretty quickly I’d need a way for the drone to recognize rooms, doorways, walls, etc and be able to estimate distances from just one camera alone, maybe using ghetto stereo, ie comparing image if you move 1s forward towards whatever you are moving towards. And to do this, I’d need a virtual dojo/training environment where my drone could learn everything. And I was building that in Unity, and then got this new cool job where I’m working 12 hours a day so that’s on the shelf. Where are we at? Obviously the military is interested in these virtual training environments and I understand they’ve invested a lot of effort. So has Nvidia with their Omniverse product: https://www.youtube.com/@NVIDIAOmniverse/playlists. Microsoft also built a tool called Airsim in 2017: https://microsoft.github.io/AirSim/

Given that we have a satellite passes every 90m (for LEO), both optical and SAR and firms, how hard is that to use to identify artillery hides + firing positions? I feel like it would be fairly straightforward to train an image recognition model on this sort of data. Are we not doing this, or is it way harder than I think?

I bet they’ve also gone through most of their good trained gunnery crews.

Yeah honestly I was thinking of Paris Hilton saying “That’s hot”

Yeah, and then your corridor immediately fills up again with all the things from outside the corridor, and you are back to square 1.

Given that everything will have chips, and conductors, some sort of way to wipe out all electronics/conductors in an area has to be an evenue of exploration. EMP, or a giant high power 2.4GHZ beam, or thunderstorm with heavy rain, or something like that. Or (in a sci-fi-ish way) as a measure of shielding, give an attacking force a thunderstorm to cover them, with heavy rain to mess with small UAVs.

We can build a lot more $1000 drones than you can buy systems to defend various targets. If one of your systems can only cover say a 1km radius, well… good luck.

And there we have it. Tens of millions vs thousands, literally 3-4 orders of magnitude asymmetry in cost, not to mention sustainment and training. And all that money won’t handle the threat flying super low, which it does. EDIT: On the other hand, this probably makes MIC twiddle their nipples.

Anything that requires an operator/sends video back to origin. I think that will last a few years still. By the end of the decade, you can imagine it’ll only phone home if sees something of interest. That I have no idea how to detect in a practical, cost-effective way!

Look up Toyan. 4 stroke menthol or gas engines for RC applications weigh 400-500g. MIT or someone related had a micro-gas-turbine that was the size of a penny in the late 90s that could power a small drone for an hour. This is an ongoing field of development. Look up “shirtbutton-sized gas turbines”.