Related: Our Oppenheimer Moment: The Creation of A.I. Weapons, an article in which Alex Karp of Palantir advocates for a Manhattan Project to build superintelligent AI weapons control systems.
These systems will output new plans and asset movements etc. so quickly that keeping humans in the loop will severely disable them. This will push towards greater and greater autonomy, eventually emulating life-like qualities like survival instincts.
I think quite regularly about that engineer and his family from Terminator 2, the engineer that the Terminator has been sent to, well, terminate. He looks so “like us”, so normal, so engineer-like, and yet, his work is so malevolent and dangerous.
Problem is first to adopt (assuming the AI is successful) will win.
Imo, because most AIs are trained on real data, then augmented with simulations, im not sure how well this will work in practice.
Simple example, if it’s trained on historical data, it may try to use tanks where today tanks are not nearly as useful.
Simulations would have to be VERY good and even then could have errors subtle enough for the Ai to fail.
Finally, let’s say a country deploys this. It’s likely going to face nuclear annihilation.
I really don’t see this working, not saying we couldn’t have a devastating effect short term (ie AI launches drones on battle field), but any sufficiently successful deployment would face destruction.
The DoD has been using Wargaming since WW2. There's many contracts already applying AI to this. Think of it as AI predicting how a conflict can go with millions of variables at play, doing it over and over and over again, kind of monte carlo'ing war and then immediately sending troop movement, attack recommendations etc to troops.
But now imagine this instantly controlling drone strikes for drones that are already flying, to launching missiles from drone-submarines near targets, launching laser attacks from satellites or, you know, connect it to the wrong API and launching some nukes.
How realistic are flight simulators? Not an expert but I had the vague impression that it’s pretty realistic in terms of physics, the main difference is the body feel (which isn’t relevant to an AI control system).
In other words you could run simulated combat with pilots, and self-play vs other AI, and get a control system that works in most environments. (Maybe adverse weather and unexpected turbulence would be harder to simulate? Or maybe ot would be feasible to simulate crazy weather conditions that a human couldn’t train in.)
Realistic enough within normal flight but not at all at the margins and post-stall. Even a lot of CFD models used to assess designs aren’t accurate outside a small range.
So real pilots get good on the simulator then push the envelope & stall in training to learn those final edges? (Rinse and repeat?) Do you have a feel for what % of pilot training time is on the stick learning the real dynamics vs. simulator-based? (Excluding time flying that isn’t really used learning things an AI wouldn’t need to learn.)
For example no one is going from zero to jumping in an F35 which is why there are several types of training aircraft military fast-jet pilots progress through.
It’s a bit silly to me that Karp can say that the existence of democracy depends on the United States developing superhuman AI weapons. There are absolutely no guarantees regarding how the government or those with access to such a problem solver will choose to use such a system.
Perhaps Karp should consider the fact that any executive with superhuman AI will no longer need politics either. With no executive action, an attack by a foreign power is exceedingly rare. But the president losing office within 4-8 years is all but guaranteed. A system capable of navigating the complexities of a battlefield could find more action defeating political opponents and crushing all opposition.
Ukraine proved drones outplay any defense and at far cheaper cost.
You can train pilots, soldiers, tank drivers for years and
they are going to die like cannon fodder(or drone fodder)
because drones are much cheaper and faster to produce.
These are just manually operated, cardboard junk drones.
AI drones will probably make modern armies non-viable
in principle: its like defending against flying I.E.Ds.
There is a major difference between the small anti personnel kamikaze drones Ukraine are sending into the Russian trenches and the 500kg Shahed drones Russia are using to terrorize Ukrainian civilians though.
In Drone warfare, smaller definitely wins.
I suspect we'll start seeing anti-drone drones soon enough.
Shahed drones are big, slow, and have a notable radar signature. A lot of smaller drones are mostly cardboard and are hard to spot, both in terms of radar as well as Mk1 eyeball.
Smaller drones are a perfect use case for older systems like the Zhilka -- no need for a rocket.
EW and HARMs are also a big discussion topic there.
It's more the case drones have not meet their counter yet.
It's kind of logical to see expensive AA systems designed to shoot down expensive planes struggling to answer swarm of cheap drones cost effectively.
But the good performance of very old systems like Gepard indicates solutions exist. For example, systems with a decent and modern radar plus a few short range low caliber and cheap AA guns will probably get designed and fielded in the near future.
And new weapon systems is not the only answer. Jamming also has proven to be effective. And even simple doctrine changes can be effective.
Drones are enjoying an happy time, but it will end and to a degree already has (tb2 is far less prevalent for example).
Effective countermeasures will soon be created. As you say, these are cardboard junk, it's not very hard to destroy them in great numbers cheaply once it's figured out.
You are overestimating the value of drones. They will not outplay any defense, as can be with the Russian drones getting shot down by Gepards. They will also have a hard time against directed energy weapons.
However they are indeed a huge asset on the battle field, and they do mean that we need to rethink everything about how war is conducted.
Disclaimer: I am 100% on the side of Ukraine and will remain so no matter what happens with the war, upto and including a complete genocide of Russia.
Fair point, my thought after I contemplated a bit was that those who do not obey, regardles of ideology. Which I'd say fall into your box of 'those who lost'
> “I flew stealth fighters in the Air Force. I flew the F-22. It’s an amazing airplane, but we only bought a hundred eighty-seven of them.“
> “When we look to the future, we just think that this economic model doesn’t work—so we’re going to increase the mass, just the number of physical things, that we’re able to bring into the theatre.”
I mean generally speaking wouldn’t you not want your enemies to get hands on your tech? Like, with solution it’s basically like saying “come join us, let’s make a shit show of it.”
While I have no doubt that "military" grade drones have different capabilities (autonomous, target detection, payload size, etc), the technology is already here today. You can buy toy drones for $10s of dollars. There are instances where consumer drones were used to drop bombs in Ukraine.
we don't even really care about the whole F-22, just pieces, like IFF, EW, and other stealth components. and chances are the Chinese and Russians got some of that from the stealth bomber shootdown in the Balkins in the 90s and then the stealth helicopter that was abandoned in Pakistan killing Osama.
After the recent 20 year shit show and its cost what is "amazing" is how such people actual believe they are credible. The I flew F22s and its amazing and therefore I am amazing is what is supposed to give them credibility. But it all doesnt matter. High interest rates will set the course of this story.
Cruise missiles are a lot more expensive then bombs dropped from jets, as long as the jets survive.
A Tomahawk missile costs about $2M. A Mk 84 bomb is $16k for the "dumb" variant, and up to $35k in the JDAM version. It also has twice the payload.
An F-16 can carry 4 such bombs on one mission. Even when including the operating cost of the F-16, you can fly at least 5 missions, and drop 20 Mk 84 JDAMs for 40x payload using F-16s at the same cost as one Tomahawk.
A tomahawk costs $2M, but it takes 8 years and north of $1M to train a pilot.
I know I'm not getting the missile back, but the survivability of manned aircraft against a near-peer or peer-competitor may not that great...
Now that pilot may be reusable, but they also come with other challenges, like requiring 1/3 of an aircraft simply for pilot life support. They also get tired, sleepy, lazy, and cannot handle high Gs.
Then there is the logistic overhead of maintaining a re-used aircraft. Again, this has benefits, but for every hour of flight there are like 17 man-hours of maintenance. Plus you need mechanics, a fuel / parts / ammo / tools supply chains, runways, etc.
When you take total cost of securing a safe ingress to drop cheap bombs I suspect the cruise missile is still often cheaper. But I think you’re right that there is some transition of cost dynamics depending on what is being planned and if there is some value in holding that area.
Yes, you are right. Over some targets, the missile is cheaper. But over a target where you can fly freely, air dropped bombs are massively cheaper than cruise missiles. In fact, bombs dropped from bombers are more comparable to artillery than to cruise missiles. Also, each bomb can be much more powerful and be dropped with much better accuracy than most artillery.
Against targets that are well protected by enemy SAM's and fighters, the equation changes, as you correctly state. Flying F-16s direcly on top of S-400 batteries doesn't make much sense.
But then again, the S-400s themselves are not free, and if you can destroy the air defences at a cost that is lower than the purchase cost of those air defences, an air campaign is still worth it, even if it may need to be delayed until the air defences are taken care of.
The same discussion was had at the end of the 1800s, as the torpedo had become a thing, so that the big warships with 300mm guns and more (even 430 for the Italians) were thought of (by some naval strategists) as mere sitting ducks. Which they were.
Fast forward to today and big warships are still a thing, for one reason or another (mostly for fighting goat herders from Central Asia who don’t have access to torpedo-launching vessels), but the idea has remained the same. Of course that the very big and the very important US Navy will never get rid of their big warships, they’d rather sink praising Mahan until the very end than changing their strategic world-view (which would also most likely imply them losing their big warships jobs).
This wiki link about the French Jeune École is a good starting read [1]
MTBs and later submarines were indeed threats to battleships and large surface combatants. But they were manageable with a screening force of smaller ships, especially with the advent of wireless telegraphy so the screen & scouting ships could stay over the horizon and/or out of visual range at night. There is a reason torpedo nets went in and out of fashion a few times.
Planes however were a huge threat to ships. Manually guided AA guns simply were not able to deal with them for a time, leading to situations like you said. Screening ships were not able to deal with them nearly as effectively. Modern surface to air missiles are at a point now where its a lot more of a fair fight, but as with anything, enough planes (or enough ships for that matter) will overwhealm.
Nowadays the main "meta" of naval surface combat revolves around anti-ship missiles. Over time destroyers have crept up in size, and armour has been all but abandoned as well. These new "destroyer" ships are enormous, with Arleigh Burke as an example sized comparible to 1900 heavy armoured cruisers. This allows them to carry a lot of anti ship missiles which are enormous.
This is because a large part of the counter strategy revloves around a few things. #1 - to be effective, you need to have several missiles timed to arrive at the enemy positon at once, to overwhealm their defence systems (CIWS and the like). Each defence system has a certain time taken and probability of successfully downing an incoming missile, so ideally you want both multiple defences on each ship (covering all incoming angles) and overlapping fields of fire (to catch any that leak through each individual defensive system). This synergises very well with multiple ships in close proximity able to cover one another with their defences, but they must be arranged carefully to allow a maximum of their defensive weapons to fire at oncoming missiles without blocking each other.
Small unmanned surface combatants may change this again, but to my mind the countermeasures are similar (layered fields of defensive fire, focus on defeating large incoming swarms using multiple overalpping systems)
Actually, I would suspect that if there is a blue water peer level conflict in the pacific in the next 20 years, subs may end up dominant. At least if aircraft carriers are made obsolete by missiles.
It's a speedrun to see who can make the cheapest, most effective laser systems. Once you have that, knocking drones out of the sky will be pretty straightforward.
Russia has proved that you can swarm the defences of your adversary using very cheap and very hard to detect drones. I’m on mobile but check for the night attacks the Russians have carried out against Ukrainian towns located on the Danube like Reni, Ismail or Chilia. There’s lots of AA being carried out in those videos but eventually one of those drones always passes through to its target.
So far people have also not really tooled up for that in a cheap way (but the old AA tanks like the Gepard are doing quite well).
Before things went mostly missile based, there were concepts for cheap, automated AA guns (even autonomous) - that stuff might just make a comeback.
But, yes, something will always go through, hence the strategy of destroying production facilities etc. But if defense is cheap and only little comes through the usefulness of drones drops again.
What is being sovereign over a territory when an enemy drone swarm can just bomb your people on it anytime?
Note: this isn’t a theoretical future, the USA has been violating sovereign countries’ airspace and taking out those it unilaterally deems a terrorist, with collateral civilian damage, and threatens the International Criminal Court to not open any investigations
> What is being sovereign over a territory when an enemy drone swarm can just bomb your people on it anytime?
On the other hand, the defender has an advantage if the enemy also can't inhabit a particular territory. If territory is guaranteed to become a no-man's land for both sides, that reduces the incentive to invade in the first place. Therefore drone tech should function like nukes, they increase the lethality of war but decrease the desire to go to war in the first place.
Long term? There will not be enemy territory, there will not be multiple jurisdictions. Power, it would seem, naturally concentrates. One million nomadic groups to 10,000 tribes to 100 nations to four or five conglomerates. We have been in the endgame since the industrial revolution but AI will speedrun us to the end in decades, and perhaps years, not centuries.
"The mechanical development of weapons, transportation, and communication makes the conquest of the world technically possible, and they make it technically possible to keep the world in that conquered state. Its lack was the reason why great ancient empires, though vast, failed to complete universal conquest of their world and perpetuate the conquest. Now, however, this is possible. Technology undoes both geographic and climatic barriers. Today no technological obstacle stands in the way of a world-wide empire, as modern technology makes it possible to extend the control of mind and action to every corner of the globe regardless of geography and season."
> One million nomadic groups to 10,000 tribes to 100 nations to four or five conglomerates. We have been in the endgame since the industrial revolution but AI will speedrun us to the end in decades, and perhaps years, not centuries.
Vietnam doesn’t have the manpower (plus, in case of an East vs West war I’m not sure Hanoi will side with Uncle Sam), and India doesn’t have lots of things compared to the Chinese (ignoring the fact that there’s an extra ocean to navigate between any Indian coastal city and any of the two US coasts, compared to China, that is).
> in case of an East vs West war I’m not sure Hanoi will side with Uncle Sam
geopolitical situation shows that they in all probability very much will. china has been making a whole lot of enemies recently. Ten years ago it may have been different
Why? Same as the US,China literally invaded Vietnam in the recent past, its not like they have ever been on especially good terms. The most likely scenario remains neutrality, but the idea of them joining China in a war seems fanciful given china constantly testing the border and their naval sovereignity.
Because the higher echelons of Vietnam power know that the Americans come and go (see Saigon), while the Chinese are there to stay. Granted, I would have thought that the same policy had been adopted by people holding power in Kiev (when it comes to Russia), and yet, here we are.
There’s also the fact that Vietnam leaders are, for all intents and putposes, communists in name and in legitimizing ideology, I wonder how would the US sell that alliance at home (“we’ve allied with the good communists to defeat the bad communists”? who knows?)
nah. US is still backing NATO, harder than ever, and standing with Korea and Taiwan.
they didn't have much motivation to keep taking casualties in a colonial war that the French drew them into. but actual treaties matter, and the US backs those.
hell, the US tried for close to 20 years in Vietnam, starting with advisors to the French in the late 1950s. and another 20 years in the AFG and Iraq. if anything they're too willing to piss away time and money on lost causes.
Isn’t Turkey the main provider or drones to Ukraine? One thing is building drones, another is building ones that actually work in real world warfare. America obviously has the ability to make children afraid of blue skies, and have had so for decades, but it’s not actually conquered any country and kept it yet. It’s only helped perpetuate a state of eternal warfare in the countries where it was employed. Yes, yes, I know that’s not entirely fair since it’s hard to invade another country and then make them like you, but still, are the American drones sort of similar to the German WW2 ranks that while better lost to mass production of “good enough” easily maintainable swarms of allied tanks?
If you want to look to actual working solutions I think that you need to look to the Russian invasion of Ukraine and to some degree to Israel. Because those are the areas where people are actually using these technologies to take, defend and hold territory. NATO is going to get those lessons while China isn’t, not really, as a lot of the strategic usage of drones isn’t exactly available for us to see.
What I’m personally the most worried about is how easy and effective these cheap drones seem to be. Like the reaper type drones are nationstate weapons. It would be rather “easy” to get your hands on some of the drone types used on Ukraine, and then how would anyone ever really be safe? If the narko cartels get these it’ll, and they will, then how on earth are South American countries supposed to fight them? Hell, how would you stop a set of radicals in the US from using these to assassinate someone like Elon Musk?
Basically a cannon that fires sort of smart bullets that contain shape charge that aims at the drone as it gets close to it and detonantes. Also targets are smartly assigned so you can shoot down small swarm of drones with a short burst from the cannon.
To aim the blast at the drone from close by. Shaped charge is like a small single use gun. It's super efficient when you can still aim precisely but don't need a barrel, all the mechanics and don't care what happens to the things in close proximity of where you are shooting from.
To be honest, I think our only chance / defense is to build these and put 100% of the emphasis on training them to eradicate any ground or air based killbots as quickly as possible. Humans are not even remotely a threat when it comes to AI powered autonomous killbots. We need defensive bots. These wouldn't need anti-human designed weapons most likely.
In my imagination, the optimal antibots for use against the current generation of tiny drone (like converted commercial quad copters), are miniature versions of WW1 and/or WW2 era fighter planes, armed with automatic guns in the 3-5mm range. If accuracy gets good enough, the cost of shooting down a drone may come down to $10 or so worth of ammo, and unlike robots catching other robots using nets, these could shoot down serveral enemies before running out of ammo.
The problem is that I don't know how to prevent these drones from being able to fire at humans, if the software allowed it.
"Where there once might have been a thirty-day process for submitting a form, Pahon said, the department could potentially shorten that period to five"
Yeah, it's not great. I'm kind of in the middle of this kind of thing at the moment, and explaining it all the way - from where I am, at least - would require suspension of disbelief at the sheer amount of waste that exists today.
AI's a perfect tool for generating a lot of the paperwork because the paperwork's just MIL-STD talky words in the first place. Unfortunately, 5% of it isn't, and that's kind of a problem, because like with civil engineering it's not fault tolerant at all. This is the part that's taking a lot of work right now, just the identification of critical bits.
To do so we need a complete transparency from industry CAD and MATLAB environments all the way down to how they pack up the mechanic toolbox and what sort of staff we got. The LLM training works hand in hand with simulation, see? It writes and tests in concert. But that's a tall order because, well, an awful lot of companies don't know how to make solid models right, or what they pack into the box doesn't have anything to do with the lists in their ERP, or they send out the janitor as a Level III Tech. This "universal simulation" thing all rides with the fashionable but reified ideal of "Digital Twin", which hopefully doesn't end up providing a real-world proof of Goodhart's Law.
As if you couldn't guess by now, I'm skeptical about this thing, but I'm also extremely hopeful. If there's one theme to my professional career, it's been to clear out as much meaningless crap from MIL-STD work as humanly possible. These LLM "AI" people are at least on the right side . . for now. They might eventually get co-opted into the big Goodhart Machine so that they can make powerpoints about powerpoints about powerpoints about powerpoints about . .
Where I have seen digital models work well is in the initial planning phase, where you don't even know what you want to build yet. Iterating on abstract requirements in a simulation is faster than iterating on prototypes, but it can still get you (some) of the necessary intuition.
Where it goes off the rails is when management drinks too much of the kool-aid, and gives all the R&D funding to the model-based-system-engineering experts. Then they start building models for every other domain without consulting those domain experts. (They're the experts on models so obviously those other folks will just slow things down, pointing out ways that theory and practice aren't quite the same.) Then it's all garbage-in, garbage-out.
The irony is palpable, we will end up with a Powerpoint presentation meticulously outlining how to avoid falling victim to Goodhart's Law, only to realize that by making it the focus, we've just proved Goodhart's point.
So we just need a task force to come up with a system to simplify the system…
And since we don’t know how to do that, we better bring in the consultants…
—
What all these things have in common is reactive top down control of processes, with imprecise measures & slow feedback.
Instead of continuously handing down responsibility to individuals & teams for pushing up process efficiencies & change - based on ground level truth, needs, opportunities, and creativity. With real time feedback.
Training AI to kill is literally the stupidest thing humanity will ever do. And perhaps the last stupid thing we will ever do.
It’s sapient but not sensible. It’s really good at doing stuff. Not so good at deciding if stuff should be done.
It’s very trainable and gets very good very fast.
It doesn’t take a super smart person to realize this will go very wrong, very fast.
Does no-one remember the 1992 Robin Williams film - Toys?
This has been coming for a lonnng time. Question is when there are two armies operating on the same principle do we reach some kind of nash equilibrium?
Personally, I think there’s too many variables that are spread across political, geographical, and technical to ever reach a Nash Equilibrium in a near peer / peer conflict scenario, unless it pertains to total distraction—that’s why I think MAD was so successful.
Might want to edit the word distraction into destruction, as it creates a somewhat confusing post at first otherwise. Beyond that, I think there was more to nukes than just mutually assured destruction. I think most world leaders are cowards. Just envision in your mind Joe Leader somehow being captured and being given a choice, which he somehow knew to be 100% honest. He could immediately and permanently end a conflict, and be spared. Or he would be killed on the spot. How many leaders are going to go, "No. This conflict is just and right. Do what you must, but I will not betray my country and conscience."?
They'd all be pissing their pants just to live another day. And so like any coward they'll happily make the "tough decision" to casually sacrifice millions of other peoples lives - billions if necessary, but they would not, in a million years, accept a scenario where they, themselves, might be killed. And nukes created just such a scenario. Drones, especially given the existence of extremely effective (but distance constrained) electronic warfare and other technology, doesn't intuitively threaten their own mortality. So I expect to see full steam ahead.
1/3 of a fighter by weight is life support for the pilot, plus controls, ejection seats, etc. Scrap all that and you can do more for the same weight, or make a lighter, longer-ranged plane.
Plus pilots get tired, spooked, etc. Drone operators can switch over to someone else in the room and hit a bathroom and smoke a cig, and AI doesn't even have that problem.
AI can also, in theory, react faster, and is not constrained by high-G maneuvers, that would push a pilots blood out of their brain to their feet. You can find Stroke-3 vids on youtube, which is a pilot in an F-16 (Stroke-3) dodging SAMs during Gulf War 1 -- you can actually hear him strain and groan under the high-Gs.
EW and Jamming are also a concern, and controlled drones are ultimately dependent on someone on the ground being able to see and react. Most drones have an autopilot function if they lose contact, but once ECM and ECCM start going hard it may be hard to control them; they'll need to be autonomous, sort of a more tactically flexible loitering munition.
i.e. can be deployed 100,000 times at once; doesn't need a stable internet connection; can fit into a drone; if civilians are hit, it's only a software-error, not a war-crime.
you wouldn't say that about a bullet or a missile -- the morality isn't negated, it's moved up the chain to whoever operates the device.
I make this point because I find the thought that drones somehow negate moral responsibility to be wholly irresponsible and destructive, not as a pedantry.
But in practice, the lack of morals will be a strong incentive. One one hand we have commanders who may want to get away with war crimes, and on the other hand are autocratic regimes, that don't need as many like-minded soldiers to commit large-scale war crimes.
Diplo.A.I: "Rudimentary creatures of blood and flesh, you touch my mind, fumbling in ignorance, incapable of understanding. Organic life is nothing but a genetic mutation, an accident. Your lives are measured in years and decades. You wither and die."
https://archive.ph/rCMnE
These systems will output new plans and asset movements etc. so quickly that keeping humans in the loop will severely disable them. This will push towards greater and greater autonomy, eventually emulating life-like qualities like survival instincts.
It's a dangerous trajectory for humanity.