Wargames and strategy require cognition and level of contextual awareness and understanding of human motivation and intent that falls well into the realm of sci-fi AGI. No military confrontation is generic enough to run on some ML pattern-matching or travelling salesman type algorithm.
I mean just to talk about how low-level and limited AI is, when you take any modern learning system, you teach it Atari Breakout, if you adjust the panel by a bunch of pixels or rotate the screen you need to relearn or transfer-learn the game because the algorithm doesn't even understand the ontology of the game and really just manipulates pixels.
Even Deepmind's starcaft bot was stupid in this sense. It didn't understand starcraft strategy at all (which was visible in it's nonsensical unit combinations it was building) but simply out-microed its opponent, it just had better mechanics.
Protoss Alphastar was fantastic. It was definitely not just out-microing opponents (after later refinements Alphastar's micro limitations were very human-like), but really had an understanding of build orders, positioning, choke points and the like. Here's a great game it played against Serral (one of if not the top Zerg player in the world) where it holds fantastically against Serral's aggression and is able to use the layout of its own base to pick apart what looked like a winning position for Serral: https://www.youtube.com/watch?v=C6qmPNyKRGw
That's something which was mentioned when it came to deep mind team talking of all the "different agents" they had trained
AlphaStar's training process consisted of two phases: an imitation-of-human-players learning phase that created a base set of "templates" and then self-play among agents seeded from those templates.
When creating those templates, they basically ended up with "baked-in" build orders. But self-play among agents ended up creating a variety of reactive strategies on top of those build orders, so that agents would modify them on-the-fly as the game situation changed. At a high level this is basically how humans use build orders as well.
Both of those (and more) are filtered by EAPM measures (effective APM) and even on those metrics Alphastar still had a lower EAPM count than most pros.
I mean looking at that Serral game I don't see any superhuman micro on the part of Alphastar. Just fantastic exploitation of base layout and steadily pressing every strategic weakness Serral made in his attack.
If anything on the consistency front, Alphastar's main advantages came from superhuman macro (effectively never forgetting about making sure workers are busy), but the edge there vs human pros is pretty slight.
> steadily pressing every strategic weakness Serral made in his attack
Again a very micro limited activity. Alphastar wasn’t very superhuman in that someone could get those inputs down about that quickly, they just can’t keep the pace up over time. Endurance is one of those human limits that is often ignored when talking about esports, but it’s real and huge influence on gameplay.
This is a very narrow view of AI. In fact, the limitations you are talking about apply only to deep learning (which is frequently and mistakenly confused for AI nowadays). Other well-studied branches of artificial intelligence include planning [1] and reasoning [2].
By far the biggest problem is to formalize the real world and build a realistic simulation which the AI system can play with. After that, a decent system would blow most humans out of the water.
> Even Deepmind's starcaft bot was stupid in this sense. It didn't understand starcraft strategy at all (which was visible in it's nonsensical unit combinations it was building) but simply out-microed its opponent, it just had better mechanics.
Maybe, what about go and chess? The feeling there was that humans did not understand the game very well.
[1] https://en.wikipedia.org/wiki/Automated_planning_and_schedul...
Deep learning is also where all the money and attention is going, and where the revolutionary gains have been seen in the last few years. They’re pretty good at compressing huge datasets into a queriable structure and generating features for hierarchical data, but are just one component of a generalized system that could handle something like war.
> By far the biggest problem is to formalize the real world and build a realistic simulation which the AI system can play with
Strong agree! The fundamental challenge is to create digital model of the world with enough detail & accuracy that you can run optimizations on it and trust the results. This is as true for robotic aircraft as it is for autonomous cars or an e-commerce data warehouse. People need to decide what subset of the world to observe, deploy & maintain sensors to collect those observations, build information systems to organize & serve the observations, and maintain the whole thing. Then maybe you can start running AI/ML on it.
It’s still not perfect. It still can’t play the full game with no restrictions. But it is discovering new strategies that humans never did. It’s not just micro.
The AI consistently did a 'deathball push' strategy, where small early advantages snowball and they pushed hard to exploit that early advantage. One of the things that enabled that was that resources were more evenly spread across the team, allowing the heroes that were strong earlier to do something with that advantage. Competitive and casual ranked play hasn't switch to always do deathball, but there is a marked increase in the amount of resources that early-timing 'support' heroes take, they are a lot greedier than before OpenAI. (as pointed out by a sibling post) It's unclear if that is either because of lessons from OpenAI-five, or because of the balance changes (there is more resources to be had around the map these days).
The second thing that stands out is that the 1v1 OpenAI and the 5v5 OpenAI did buy a lot more of 'consumable' items, mana and HP potions basically. These were considered too inefficient before, people tended to prefer items that gave a more over-time regeneration and long-term efficiency. What the bots did was /constantly/ ship consumables over, which was kind of surprising at the time. It used that consumable resource to either force an opponent away from xp/gold resources (lane, jungle) or allow it to stay after aggression. Once pros started buying a lot more consumables, casual players quickly followed suit.
FWIW the developer of the game has also tried to fix this over the years, devising ways to get more resources to the 4th and 5th players. For example, there is a system similar to UBI where all members of the team benefit equally from "bounty runes". But till date, the bots have achieved a level of egalitarianism yet to be matched by human players.
this kinda makes sense. I'm more of a cs player, so hopefully my observations aren't totally off base. generally, a group of five humans will have some sense of who their best player is, who their worst player is, and who's good at specific roles. for instance, you might deliberately send your worst player into a bombsite first to gain info about the opponent's defense (and save your best player to react to that info). or you might send the person who's best at "entry fragging".
an egalitarian playstyle doesn't make sense when you have human players with different capabilities. you want to play to each person's strengths. if you have an algorithm controlling five heroes (or five instances of the same algorithm, idk how it works), there's no "best player" to send more resources to or dedicate to a specific role. I don't know enough about moba to say whether this is even an advantage.
none of this stuff is particularly interesting in counterstrike. a circa 1999 aimbot with a good ping will beat any human player.
This gave him an extraordinary ability to contrast live battle situations with a huge library of learned historical situations, helping him to make decisions on battle plans.
Probably because you’d end up with an AI that would do something that fulfilled its training criteria, but didn’t really reflect the real desired outcome, like that Tetris AI that just paused the game when it realised it was going to lose. Or the AI would suggest attacking unacceptable non-military targets, or inflict collateral damage, or suggest nuking the enemy as a first strike. AI might be an interesting tool for wargaming, but it’s never going to be a complete strategy guide.
I’d say that AI is used extensively throughout the defence industry in specific areas though, and I’m fairly confident wargaming is one of them. But there are people who have context the AI will lack at all points in the chain. You wouldn’t just have some AI ordering troops about.
I would expect the people who wrote the AI would impose on it constraints to avoid this – to be useful in the real world, an AI has to comply with the same constraints that real world military commanders have to follow, and the laws of war are among those (certainly if we are talking about the US military.)
(A realistic risk is that it might interpret those rules differently than a human would – maybe in an overly pedantic manner – and its overly pedantic readings of the rules might give human commanders cover or encouragement to engage in the same harmful pedantry.)
> or suggest nuking the enemy as a first strike
I doubt they'd even give an AI that option to consider. Realistically if you are a US military commander – in Afghanistan or Iraq or Syria or wherever the US might intervene next – that's not an option on the menu. The decision to use nuclear weapons only comes from the very top, and the very top will only consider it in utterly dire circumstances.
To be actually a useful tool to military commanders, you want an AI which suggests to them strategies they might not have thought of, or which gives them a reasoned opinion on which of the various options they are considering is most advisable. A useful AI tells them how to use the resources they have, and nuclear weapons are not–in normal circumstances–an available resource to US military operations.
The devil is in the details.
What if the AI decides that the course of action that leads to the desired outcome of winning the conflict is to kill civilians but it is also aware that outright killing civilians is against the law or will be ignored by the decision making human so it instead decides to recommend a series of strikes that indirectly lead to a large famine which results in the death of the civilians and winning the conflict?
Who is to blame there? Is this outcome desirable?
"To be actually a useful tool to military commanders, you want an AI which [...] gives them a reasoned opinion on which of the various options they are considering is most advisable."
With that (far more sophisticated) AI system, leadership could just de-prioritize options with excessive risk of civilian casualty.
So I don't see why a strategic military advice platform couldn't do the same basic thing.
[0] Here's a working example: http://beta.datalex.org/app/?rulebase=http://austlii.communi...
What if the AI lies?
What if the AI lies and civilians die because of it?
I always figured that was what happened in the film Colossus: The Forbin Project, the computer(s) found a way to ensure / prevent a nuclear war / peace, by taking control of the nuclear arsenals on their own and ensuring peace via their own threat of nuclear weapons.
Requires political victory. War is extension of politics, not alien invasion defense.
Thus, a planning AI that can handle tactical targeting task but can’t handle politics probably “can’t war” going by definitions.
1) A military commander programs an AI squadron with the parameters of destroying a heavily defended military target with the parameters of minimizing casualties.
2) The AI veers the bombers 180 degrees off course and attacks a nearby town, purposely and irreparably destroying the only hospitals and drinking water production facilities in the region. Thousands of civilians are killed, and the loss of fresh water will cause many more casualties in the following weeks as people are no longer able to get drinking water. No military targets are hit.
3) The AI is examined. The reason it attacked the town is unclear but it is determined that it was not a malfunction or sabotage.
Given that.... was a war crime actually committed?
The AI cannot commit a crime, it's not a person. The people that ordered the attack specified that it should be attacking a valid military target. The people that created the AI certainly didn't intentionally put any mechanism in there to have the AI purposely commit war crimes.
Given that, is it an actual crime? Or would it be closer to something like a natural disaster?
...to be clear I'm not a bright guy so I've probably missed something obvious...
https://gizmodo.com/for-20-years-the-nuclear-launch-code-at-...
As a preparation activity you want to be prepped for both perfect play opponents and imperfect opponents that do crazy things like nuke first. Arguably such a system would also help eliminate bias in (pretend) adversarial commanders.
Never is a strong word.
Just like Goalkeeper will beat any human operated gun, a cheap disposable drone swarm pointed at an enemy country with a program to blow up anything resembling an airfield will beat any military general at leading an air campaign, and will probably beat even a nuclear weapon in effectiveness. It's already possible without any real AI. It's just nobody bothered to make one (or probably did, and is keeping quiet about it.)
It's just we did not have a big war in the age of powerful, disposable computers, and thus never weaponized it. There are many weapons today which claim to "beat nuclear weapons" in their respective context, with the main determinant of them being "super" weapons is the lack of humans in the loop.
All of them can be developed, and weaponized, it's just a man with a gun, artillery, and amassed air, or armour already works superbly well at the task, and most importantly does so reasonably economically.
I don't think that exists. We have cheap drones that won't do much to an airfield and don't have much range.
We have B52s that can blow up an airfield but those aren't drones. Or cheap.
The Earth is too small and our weapons too powerful for playing child games anymore.
So it doesn't really make sense to train an AI for full scale conflict, nuclear arsenals are already MAD-sufficient enough. Instead, you'd have to model complex geopolitical, even cultural and psychological details to model surrenders and which territories one could lose, for example. Essentially AGI-complete tasks.
I know our realities are such that we can't let go of defense spending and defense systems yet, but it's looking increasingly ridiculous and dangerous to rely on military strength[1].
[1]: Instead of seeking ethical-economic-cultural multinational equilibrium, that is, a mutually acceptable plan toward individuals and development of our entire society/civilization.
In the sense of each "nation" (non-trivial political subdivisions in general) tolerating the collective vision for development of our society. I don't think bombs are a very effective form of communication or cultural-philosophical exchange.
Conflict will increasingly be drone based. The controllers of those drones will increasingly be wielded by AI. This link is about tactics. This thread got derailed into strategy. But it’s reasonable to think that AI would do a good job of controlling attack air/sea drones with clear success criteria.
The standing armies are generally smaller than the ones before and after WW2, despite three times larger worldwide population. The effective mobilization potential of major countries is much smaller than before or after WW2 as many countries have moved on from total conscription where you'd expect most of adult men to have military training and a huge proportion of the population to actually fight in a war. The quantity of planes, bombs, guns and shells that we can muster is an order of magnitude smaller than 70 years ago, so while each bomb or gun is more effective, the total destructive power is less than it used to be. The world has 5 times less nuclear weapons as it used to have.
So the actual trend is towards a smaller scale of destruction. We have major wars that result in as much casualties as an average WW1 tuesday despite being as full-blown conflicts by the involved countries, mobilizing as much as they reasonably can in modern circumstances.
All of these cases assume you have simulators that represent reality realistic enough that you can use AI approaches to find ideal solutions in the parameter spaces. Dog fights make sense since there's a lot of computational work done in the aviation industry and the environments are pretty stable, as I imagine constraints on aircrafts and pilots are probably pretty well known. Other cases may be a lot more messy to accurately model.
Exactly. This reminds me of the following quote:
"AI is whatever hasn't been done yet." -Larry Teslar
https://en.wikipedia.org/wiki/AI_effect
AI is used all over the place in the military, but we sort of ignore the boring AI used e.g. in logistics in most conversations.
1: Having ai wargame it out ala deepzero would require modelling war accurately, which is a nigh impossible challenge in the first place.
2: Using past battles as data for generating tactics is useless as each successive generation of tech redefine the battlefield.
A more fundamental challenge is about hunches the generals have about what they don't know due to always incomplete intelligence data.
One notorious example was the US Air Force's Advanced Penetration Model, which due to a programming error made US bombers invulnerable to enemy air defences by inadvertently altering their latitude or longitude when checking their location for a missile impact. This had the effect of 'teleporting' the bomber, at the instant of impact, hundreds or even thousands of miles away, causing the missile to miss. Furthermore, this error went unnoticed for a number of years. Other unrealistic models have had battleships consistently steaming at seventy knots (twice their top speed), an entire tank army halted by a border police detachment, and attrition levels 50% higher than the numbers each force began with.
If you can’t validate it, do you dare trust it above humans?
With tactics, validation still is hard, but you can do (expensive) real-life experiments.
Ideally it also has to work in a grander geopolitical context, which is a lot of complexity to account for and simulate, involving a lot of very abstract and subjective calls.
So it's probably a much more complicated problem than it might appear at first glance.
As an example - if we had modern ai in ww2, it would help with the position of tanks and soldiers, but wouldn’t invent the idea of breaking Enigma, or the idea of making decoy tanks for the d-day.
I suspect if a serious war started one side would just lose, almost as if by accident. I have no reason to believe this.
NB: self-driving cars, drones, and Boston Dynamics robots are all examples of systems that make significant use of hand-designed control algorithms. AFAIK none of those systems are doing (especially end-to-end) deep rl for control.
I think the answer to your question might be something like "neural nets aren't as good as you think and have gotten a lot of credit for things they aren't used for and aren't even very good at."
AlphaStar: excuse me??
People write that these games have predefined rules, as opposed to real-life scenarios.
Such people have probably never actually played Starcraft or any other similar games. Sure, the rules of how much damage each unit does, etc. - are fixed. (Actually not even they are fixed throughout the months, and a good player needs to be able to adapt to changes). But what the actual games end up being rarely depends on these rules directly.
What happens instead is a meta-game that is created between you and your opponent where you have to read their strategy and come up with creative ways to deal with it.
In fact in real life the situation is similar: the laws of physics are also always the same, the stats of airplanes are also the same for most of a given year, how much damage can it do, which angles of attack it can support, its speed etc. But it's the combination of these rules in real life situation that makes for the real meta-game.
AlphaStar is absolutely an example of an AI that deals with this exact problem.
PS. Sometimes I see how in the future when we create human-capable AI, these people will still say "but it's just statistics, it's not real intelligence, it's just advanced SVMs". Who cares about what the underlying primitives are? 2+2 does not equal 4 in complex systems like this, it's the emergent endogenous effects that produce all the bang. I don't care if it's all 1s and 0s, or some other small statistical primitive that you could reduce everything to: what's importan is the high-level computation and what it can achieve.
No soldiers to lose, no resistance to wars and use of weapons. Because we don't care about the rest of the world that we drop bombs on, that's for sure.
Robots don't say no, they do not need indoctrination, they keep secrets, they are expendable, they don't receive benefits, it doesn't take 20 years to make 1 robotic soldier at a time, they're repairable.
If a country decides not to go robotic, other countries will, and those countries will be in control.
Weapons systems operators, transmissions operators,...
The goal never was to have boots on the ground, but rather to have guns that shoot where you want them to, when you want them to.
It just so happens that until now that required a fragile meatbag in fatigues & boots to pull the trigger.
It'll probably be so effective, it'll still be fighting long after all of humanity has destroyed itself and left an irradiated environment.
Here are two science fiction short videos of an artist depicting such a future:
"LAST DAY OF WAR" - by Dima Fedotov | TheCGBros - https://youtu.be/IjJmTeBSEzU
"Fortress/Крепость" - by Dima Fedotof | TheCGBros - https://youtu.be/pyMNIFZTQkg
(edit: typofix)
AI actually gets into trouble due to the amount of data before humans do. Humans are quicker learners and only need a few iterations of analysis, while current AI needs a plethora of iterations over possible combinations.
For example the OpenAI Dota 2 AI was doing pretty well, but it had to play a restricted version of the game where only about a dozen heroes are allowed out of ~120. Basically only 10% of the complexity. OpenAI never even attempted to compete at the full game before they quit and shut down the project.
When a self-driving car company can have its AI drive around in an Arizona suburb, nobody sane believes that it means they can also drive around in Kolkata, India. However that is basically what happened with the Dota 2 AI effort by OpenAI. They managed to succeed in an extremely limited scenario and then proceeded to put out press releases claiming they had beat Dota 2.
It's difficult to admit that what tends to make you a smart and valuable human is your ability to behave more like a computer. It's even more difficult to admit that the parts of your job you consider the most fulfilling and impressive can often be rendered unnecessary with minor process changes or ought to be done by someone else whose skills are more appropriate for those tasks.
You can build simulations and train it on that but then it would find some quirks in your simulations and exploit that or your simulation is too basic and won't model some variables that is relevant and a human can exploit.
Then if you have the perfect simulation with all the rules and all the variables you should probably use computer science and find the best strategy and not some statistics/guessing/AI stuff that might find a good enough solution.
I think you don't have enough real world data to train an AI, probably you could put the effort in building an AI to solve the economic and social problems, find better way to solve the issue where some guy spends on Fortnite skins the money that could feed an entire family... I would try to create a simulation for the economy, improving the lives of the poor around the world should reduce the political tensions and reduce the number of conflicts IMO.
You're never going to hear anything about the details of how this works or whether or not it works at all because it only works if it's kept a secret.
Something like speeding up TCPED seems like a much more modest and immediately doable goal; reminds me of efforts to automate financial analysis to some degree. Love the codename if it's actually genuine.
And yes, obviously secrecy makes sense, but I'm not sure to what extent. Obviously the whole point of MAD was that it was loudly trumpeted as a deterrent. It doesn't seem impossible that the same might happen here. On the other hand, the military certainly has some ability to keep things quiet when they need to.
That's what these F-16s are trying to do. It is taking visual and sensor information by the truckloads and find patterns in it, make meaningful conclusions from it and provide direction.
Your definition of AI is too narrowly scoped. AI can be applied in planning, strategy, logistics, and things like flight mechanics, as well. This is precisely because AI today itself is narrowly focused. We don't have AGI obviously.
Yes, and no. If we break down by time horizon, this is the feasible limit: solving constraints and at-risk problems for operations planners; solving relaxed instances of future scenarios to provide guidance to mission planners given partially observed opponent behavior; and engagement level models. Only the last level is really something you can generate / extract training data for, and at vast effort. We've been modelling air combat for 40 years or more and it's still difficult for larger scenarios. We really have only a nascent concept of how to employ unmanned jets.
Amateurs study tactics, professionals study logistics.
You don't want AIs replacing human generals, because generals need to make tough calls. So tough, that their decisions are still debated by us 100+ years after. War is the most sophisticated application of human intelligence. Dog-fighting F-16s are more like a self driving car in this sense, than a general intelligence capable of winning a war.
AI wrote the logistics plan for Gulf War 1 and more than paid for itself.
This is a great way for optimising towards winning past conflicts. It is a terrible way to adapt to new developments.
A very good example would be the mass employment of "cheap" drones in the recent Armenia / Azerbaijan conflict.
There are enough records on the drive itself about how the Military has funded multiple supercomputers and algorithms including popular videogame engines for this very purpose.
Strange game, the only winning move is not to play.
Yes, I realize I'm almost literally describing SkyNet, but isn't that where the power would be, for better and worse? The strength of computers has always been the ability to analyze literally incomprehensible (to humans) amounts of data and derive unique insights from them, and current machine learning has magnified this ability several-fold. AIs have now repeatedly been dropped into all kinds of video games and gone from morons to competent to beating all human players and exploiting bugs no one even knew existed. In a grim sense war is just another game. It has constraints, resources, definable win and loss states, etc. Consider that conflicts have often tilted based on a single novel strategy discovered by limited human commanders, and then ponder what an AI that improves its Mario speedrun via nearly imperceptible collision glitches could do.
Potentially, weapons don't matter as much as how well you can use them.