Discussion: AI and the future of DCS

[TheFreshPrince]

I would like to kickstart a discussion on how you people but also the devs and ED team see the short term future within the next 1-2 years regarding the insanely fast development of AI. Which can already be used pretty well for coding with Chat gpt 4.0 after what seems like only a few months of development. Imagine what it would be able to do in a year if we keep up this pace...

To me it seems like this opens up possibilities of speeding up the development of DCS so quickly within the next few years that it might be possible to put together everything that we would want to have in DCS. 

But I'm not a dev and only have basic programming knowledge, so I would like to have opinions from you people and discuss where this could potentially bring us to. 

[Mars Exulte]

Funny, I mostly hear how ChatGPT sucks for coding because it makes up complete nonsense. Even if it didn't, its ability to code for a proprietary closed eco-system with no publicly available documentation is exactly ZERO. It EMULATES stuff it encounters online, and may or may not actually do anything.

There is no magic ''mek gam fastah plz'' button, and unlikely to be one anytime soon. AI is a very useful tool, in limited work case scenarios. It is not a magic button. Any articles you read written in breathless terms about how revolutionary the tech is, how it'll change the world any day now, blah blah, are about as credible as all the crypto garbage from a few years ago (ie written by somebody trying to bait investors).

It is a tool. It will find itself useful for certain niches and not for others, just like all the other tools ever developed in human history.

Edited May 4, 2023 by Mars Exulte

[MiG21bisFishbedL]

2 hours ago, Mars Exulte said:

written by somebody trying to bait investors

 

 

 

This is the real take away. It's mostly hype and effort to bilk investors.

The old programming maxim of garbage in - garbage out applies and since AI needs to sample existing works? I can't see its value being much more than something a developer can use to get rid of some of the repetitious work of coding. It's a smart Ctrl+V in a sense.

But, it's no magic button. Unless, we're talking a magic button that you can press and run the risk of getting a lawsuit if you're not sampling your own code work.

Edited May 4, 2023 by MiG21bisFishbedL

[Dragon1-1]

Yup, this is simply investor hype. For all the big words written about it, AI is mostly good for producing good-sounding prose. If what it wrote is true, it's only by accident. That's not even getting into all the copyright problems that are about to come down on it (turns out, IP rights and licenses don't disappear just because something is part of "big data").

It's not even the first time machine-generated text was used to fool a human:
https://en.wikipedia.org/wiki/SCIgen
2005, much less advanced than ChatGPT, and admittedly single-purpose, but it worked on people who didn't observe due diligence when reviewing the papers. Truth to be told, superficial human interactions are easy to imitate. If you get down in the weeds, particularly with something that requires any sort of context, AI will often turn out total BS. There are jobs that it can replace, but they could probably have been replaced by a small bash script, their holders were just able to dodge that bullet by a combination of luck, politics and their own BSing powers (also helps that most people have no idea what computers can already do). In the end, for all its computing power, ChatGPT is nothing more than an autocorrect writ large. It likewise has no context and makes exactly the same kind of errors.

[Mars Exulte]

1 hour ago, Dragon1-1 said:

Truth to be told, superficial human interactions are easy to imitate. 

 Back in the 90s my dad had a BBS set up using ShamPage as an automated help system. It was purely exotic scripting triggering off keywords, with some ''imitate a human typing'' stuff like hesitation, making and then correcting spelling mistakes, very basic stuff. If you asked questions about the board, it could answer them, in detail, as he had carefully programmed it for that use case.

 

But if you were rude or demanding, it would trigger off that, too, and quickly escalate to antagonistic insults. Again, just scripts and keywords pulling from a database of prepared responses. The number of people who would argue with it in circles for lengthy periods was truly remarkable, and there was no AI even involved. It's easy to fool people. Just look at how many run around copy pasting sales brochures around here like it's Jesus' Second Coming.

Edited May 4, 2023 by Mars Exulte

[TheFreshPrince]

So, if AI was that bad as you guys describe it, why are top techs and MIT professors already calling for the halt of development of future AI?

https://futureoflife.org/open-letter/pause-giant-ai-experiments/

I mean, if your own devs are quitting, because they get scared of what they created, you know <profanity> is getting real:

https://www.theguardian.com/technology/2023/may/02/geoffrey-hinton-godfather-of-ai-quits-google-warns-dangers-of-machine-learning

I think you guys are underestimating the capabilities of current AI (GPT 4.0 etc.), which is only in a very early stage of development. Look up for yourself what Midjourney can do for example. Or GPT 4.0. It's open use, so it has beend tested and verified by literally anyone on youtube. This is not investors hype BS, there are no investors needed. Those AI are from the biggest tech companies of the world and they are, again, open use. And they are, in contrast to NFT/crypto bs, usable and being used already.

Older versions of ChatGPT or older AI is not even worth to be mentioned in this discussion, of course, since it was garbage. I'm talking about the capabilities that GPT 4.0 already provides (said to be 100x better than the previous version). But also the potential that future versions and other AI stuff will bring in a short time span. And yes, programmers/devs will still be needed to implement this. But it's simple math. If you can programm 10x faster with the help of AI, you can publish stuff that would take 10 years of development within only one year. That would be probably anything from any wish list that people have created on this forum...

But let's ask ChatGPT on this topic. Here's the answer:

Zitat

ChatGPT 4.0 can significantly contribute to the development and improvement of DCS:World, a digital combat simulator, in various ways. While it may not be possible to quantify the exact amount by which future AI versions of ChatGPT can speed up the development process, the potential impact can be substantial. Some ways in which ChatGPT can contribute include:

1. Design assistance: ChatGPT can help create detailed design documents by generating ideas for new aircraft, environments, and missions based on existing data and user requirements.

2. Code generation: ChatGPT can provide code snippets or entire sections of code for specific tasks or functions, which can reduce the time developers spend on writing code.

3. Bug detection and fixes: ChatGPT can analyze code and identify potential bugs or areas for improvement, offering suggestions for fixing these issues.

4. Documentation: ChatGPT can assist in generating well-structured and comprehensive documentation, including user manuals, technical specifications, and tutorials.

5. Community management: ChatGPT can help manage online communities, answer questions from users, and gather valuable feedback to inform future development.

6. Marketing and promotion: ChatGPT can generate engaging promotional materials, such as blog posts, social media updates, and press releases, to raise awareness of the simulator.

7. Training and onboarding: ChatGPT can create training materials and interactive tutorials that can help new users learn the intricacies of the simulator quickly and effectively.

As ChatGPT continues to evolve, its capabilities will likely expand, potentially leading to even greater efficiencies in the development process. While it's difficult to predict exactly how much faster future AI versions of ChatGPT can make the development process, it's reasonable to expect that advancements in AI technology will lead to significant improvements in various aspects of the development pipeline, ultimately speeding up the overall process.

 

Edited May 4, 2023 by TheFreshPrince

[Dragon1-1]

5 hours ago, TheFreshPrince said:

So, if AI was that bad as you guys describe it, why are top techs and MIT professors already calling for the halt of development of future AI?

I'm pretty sure all that it is all way overblown, driven by worst case scenarios and unreasonable expectations of technology can do (tech people are really prone to the latter, even professors). Well, either that, or they are concerned about AI going along the same path as internet - great freedom, but at the price of being used to spread loads and loads of BS and outright propaganda, on top of other problems. Spammers are already milking ChatGPT for all it's worth.

AI has potential, but it won't become self-aware, gain imagination or become anything more than a deterministic logic machine. It's a useful tool for those who know how to use it. For the rest, it's a fancy toy. Destruction, or even significant disruption of society due to it is not likely, all it will cause is a shift in the job market. When you think about it, the internet wasn't a revolution, either, but an evolution. It took many years for it to become a basic necessity, as opposed to a tool for scientific cooperation and a toy for geeks. AI is likely to be similar.

Look at where the internet ended up. It's the best indication that total, unrestricted freedom is not good. Sure, it's great as it is, but the very fact computer viruses exist comes from the fact it's not inherently secure. It could have been designed without those flaws, had it been actually been designed, as opposed to being grown piecemeal by everyone from corporations to random programmers with too much time on their hands. It could have been designed so that darknet could not exist, but it wasn't, and as a result, it also serves a whole lot of criminals. To me, the greatest risk of AI is not AI itself, but people who would misuse and abuse it.

Edited May 4, 2023 by Dragon1-1

[Mars Exulte]

7 hours ago, TheFreshPrince said:

So, if AI was that bad as you guys describe it, why are top techs and MIT professors already calling for the halt of development of future AI?

 I saw when he quit, and it's all a bit overblown. Eventually, AI will become a big deal, it arguably is already, and as mentioned above, like the internet, it will be both really useful and really dangerous. You asked how AI will impact the development of DCS World. It will not in the forseeable future, and at no point will it suddenly conjure <profanity> out of thin air and make all your assorted dreams come true. As mentioned earlier, it also does not bypass all the legal issues that must still be resolved. It is a TOOL. It is not magic.

7 hours ago, TheFreshPrince said:

I think you guys are underestimating the capabilities of current AI (GPT 4.0 etc.), which is only in a very early stage of development.

 No, we know exactly what its capabilities are, which we have already outlined are currently lacking. It can generate some functional code, it must be reviewed and made sure it ACTUALLY works and be repeatedly tweaked.

7 hours ago, TheFreshPrince said:

Look up for yourself what Midjourney can do for example. Or GPT 4.0. It's open use

 Yeah, we know. I played with it a bit myself. 

7 hours ago, TheFreshPrince said:

so it has been tested and verified by literally anyone on youtube.

 Oh well, Jesus, argument over. The Tube has spoken @@ Again, I will refer to crypto, a technology that has useful use cases for our future but was grossly over sold.

7 hours ago, TheFreshPrince said:

This is not investors hype BS, there are no investors needed.

 Development costs money, and products need customers. Yes, they are hyping their products, and yes they are nosing around for money. These are commercial enterprises, dude. They literally are working on subscription models and tiers for these things.

7 hours ago, TheFreshPrince said:

Those AI are from the biggest tech companies of the world and they are, again, open use.

 For now, and not unrestricted. So, just like myriad other examples of freeware.

7 hours ago, TheFreshPrince said:

And they are, in contrast to NFT/crypto bs, usable and being used already.

 Crypto is usable, being used, and pioneered numerous concepts that will be gradually integrated throughout digital life. It was not useless, it was OVERHYPED.

7 hours ago, TheFreshPrince said:

Older versions of ChatGPT or older AI is not even worth to be mentioned in this discussion, of course, since it was garbage. I'm talking about the capabilities that GPT 4.0 already provides (said to be 100x better than the previous version).

 When the baseline is ''useless <profanity>'' being ''100x better'' is not as stellar as you make it sound. Again, we've read the same stuff you have. We've also read what was said by people not attempting to monetise an article or YouTube channel.

7 hours ago, TheFreshPrince said:

But also the potential that future versions and other AI stuff will bring in a short time span. And yes, programmers/devs will still be needed to implement this.

 Yes, we know. Which is what we said. It is not a magic button.

7 hours ago, TheFreshPrince said:

But it's simple math.

 *rolls eyes*

7 hours ago, TheFreshPrince said:

If you can program 10x faster with the help of AI

 It may be that useful someday, but not anytime soon.

7 hours ago, TheFreshPrince said:

you can publish stuff that would take 10 years of development within only one year.

 That's not likely to be how that works, but sure. If it makes you happy.

7 hours ago, TheFreshPrince said:

That would be probably anything from any wish list that people have created on this forum...

 Ignoring the fact it still needs access to detailed data about aircraft, specific physics, methods of operation, and previously mentioned legal issues, among other ''terms and conditions'', all in a digital form that it can actually parse and correctly interpret.

7 hours ago, TheFreshPrince said:

But let's ask ChatGPT on this topic. Here's the answer:

 That is called ''word salad'' and is an excellent example of what we're talking about. It has no damn clue what DCS is or whether it can actually do any of that, it's just reciting generic ideas about its potential to code without recognising any of the hurdles in front of it or whether it's ACTUALLY able to code in this PROPRIETARY engine it has not been trained on or exposed to at all, and again, has exactly nothing to do with its access to or ability to understand the data necessary to code for DCS.

19 hours ago, TheFreshPrince said:

I'm not a dev and only have basic programming knowledge

 We noticed.

19 hours ago, TheFreshPrince said:

I would like to have opinions from you people and discuss where this could potentially bring us to. 

I'm really excited about these ads and YouTube videos I saw and really don't like people poking holes in my imagination

Ftfy

[MiG21bisFishbedL]

16 hours ago, TheFreshPrince said:

Look up for yourself what Midjourney can do for example.

Make its owners a target in a class action lawsuit?

[Dragon1-1]

Yeah, we're going to be seeing a lot of those. Right now, AI training data is based in large part on unlicensed public content. If laws are passed to protect artists from their work being exploited that way, training AI would become much harder. Not to mention existing licensing issues, for example, I'm pretty sure GPL code was used in training ChatGPT. Does that mean all code it produces is automatically GPL? 

That said, I can perhaps see one area where it could be of use to DCS: 3D modeling and texturing. If someone would make an "ultra-photogrammetry" AI that could create a functional 3D mesh form a collection of 2D images, perhaps this could be used to make AI models for DCS. That said, I haven't yet seen an AI that would work with 3D models, and to create, say, a plausible looking tank model from a bunch of photos is a much more complex task than drawing a picture. Also, such an AI would require a large repository of 3D models, unlike 2D graphics and text, such repositories are not only much less common, they're more tightly locked down (meaning, you have to pay for access). This is an area where AI could probably make a lot of difference for game development, turning 2D artwork into textured 3D models is a fairly labor-intensive process.

15 hours ago, Mars Exulte said:

 Crypto is usable, being used, and pioneered numerous concepts that will be gradually integrated throughout digital life. It was not useless, it was OVERHYPED.

Crypto, specifically blockchain, is a great big dud, for the same reason AI might still be: humongous infrastructure requirements. The only reason crypto is used for anything is that markets work on the basis of a self-fulfilling prophecy - people believe it has value, so it does. This also makes crypto a highly unstable and risky investment, prone to sudden collapses, particularly when the wider financial markets are in trouble. Blockchain itself is one of the most overhyped technologies, with some outright cultish advocates, yet it doesn't actually solve any high impact problems. Proof of stake might have some future, but proof of work is just too inefficient to be practical.

Incidentally, one of the major users of crypto are, wait for it... criminals. Because, guess what, unregulated, difficult to trace currency is great for scams, ransoms and other situations where you need to move funds around without the authorities finding out. Sure, some legitimate dissidents also use that anonymity, but as it is, I believe it's ultimately doing more harm than good.

[BitMaster]

 

It's a dangerous endeavor to create a thing that you do not completely understand how it works inside

to help you solve problems you can't completely understand and therefor not solve yet.

The danger is the boundary between knowing and NOT knowing that humans cross without noticing ( Ignorance fails to recognize itself )

You will not be able to fully qualify the answer(s) given by AI as you never understood the full thing first place but you will qualify it without being qualified/knowledgeable !

There is the danger.

 

 

 

 

 

 

 

[AngleOff66]

Asimov's 3 rules of Robotics.  Change Robotics to AI.

Did ANY of the companies developing AI program these in?  When there is money to be made? What am I thinking.

 

[Dragon1-1]

Those are for thinking robots. Again, AI doesn't think. It's an elaborate algorithm that strings words or pixels together from a heap of "publicly available" (read: stolen) data. There's nothing magic, intelligent or even particularly dangerous about it. Most jobs it'll kill off shouldn't have existed in first place, anyway. AI's biggest weakness, but also the reason it can work as well as it does, is that words have no meaning to it - it basically stings them together based on how likely it is that a given word will appear after another word. It does not have any kind of context, and it'd take another revolution to give it that capability (and even then, it wouldn't actually be thinking).

32 minutes ago, BitMaster said:

You will not be able to fully qualify the answer(s) given by AI as you never understood the full thing first place but you will qualify it without being qualified/knowledgeable !

There is the danger.

Humbug. Here's the thing: AI answers questions that we ask, based on the sum of answers that were given to such questions before. It doesn't reason, it doesn't understand, it just makes educated guesses about which answer to parrot. In fact, once knowledge and especially understanding enters the picture, AI-generated BS tends to fall apart quickly. Complex programs written by AI almost never work, and if you ask it about a verifiable, but obscure fact, it'll feed you outright BS that only sounds correct. Give it a try, ask it about any obscure thing you know about.

Yes, the ignorant will ask and form opinions without it, but when had it stopped them before? Various gurus, scammers and/or politicians have made a huge load of money out of telling people things that sound plausible, but require a degree of knowledge to verify, which said people do not have. AI makes it a little easier, but that's it.

[Exorcet]

AI is something that absolutely 100% of game developers need to look into starting now, but there is still a road ahead of us before we arrive at AI generated games.

I'm not sure that 1-2 years is enough to expect anything major. Maybe in 10 years, but I'm just making guesses anyway.

The possibilities are certainly enticing, as AI could make development much faster and less bug prone. I'd also like to see it replace the traditional AI of DCS. All of this is going to take work to accomplish though.

[Dragon1-1]

There's very little AI can do for games. It can't code, it can't make textures (since it doesn't have a concept of an UV map) and it can't do anything really useful to a game dev. Maybe it could help with advertising.

I don't think we'll get there in 10 years. Not without a revolution. Most "possibilities" of AI had been dreamed up by its pushers to find investors. It has certain applications, but right now, its fundamental concept puts limits on its use. It can't handle the kind of complexity even a simple game involves, there are too many interconnected variables, and to resolve that, one needs a human brain.

Just now, Exorcet said:

I'd also like to see it replace the traditional AI of DCS.

This is not going to work. How are you going to train it? Tracks from players and the like are just too few, DCS has no "big data" to train an AI on. There was a dogfighting AI some time ago that defeated a real fighter pilot in a sim, but that was for a single plane, in a single, simple set of conditions. An AI of that kind so complex as to pretend to be a human pilot is simply not possible within ED's budget. Better to make an algorithm that would make decisions similar to a human pilot, who mostly flies by the book. Techniques and procedures that it would follow are detailed in public documents. This kind of AI is how ED currently does it, it just needs some improvements.

[Exorcet]

4 minutes ago, Dragon1-1 said:

There's very little AI can do for games. It can't code, it can't make textures (since it doesn't have a concept of an UV map) and it can't do anything really useful to a game dev. Maybe it could help with advertising.

I disagree, coding is well within the reach for AI. It's going to take time to develop AI that is good at it, but it's not a far fetched idea.

4 minutes ago, Dragon1-1 said:

I don't think we'll get there in 10 years. Not without a revolution. Most "possibilities" of AI had been dreamed up by its pushers to find investors. It has certain applications, but right now, its fundamental concept puts limits on its use. It can't handle the kind of complexity even a simple game involves, there are too many interconnected variables, and to resolve that, one needs a human brain.

The brain and AI aren't much different. The only problem with current AI is that it's a very young technology. It's like saying cars will never go 200 mph because the the Motorwagon couldn't break 20. I work with engineers and software developers who are the ones pushing AI for product development, not investors. We already have automated processes for design in fields like aerodynamics. AI can easily (conceptually, creating the AI takes work obviously) work its way into these applications to improve them even further.

 

4 minutes ago, Dragon1-1 said:

This is not going to work. How are you going to train it?

There are many methods. Tracks are a good one, I don't see why you're saying there are too few. We can create more easily. Also as AI works its way into other fields, there will probably be publicly available resources like papers on how to train AI to fly, or strategize in general.

4 minutes ago, Dragon1-1 said:

Better to make an algorithm that would make decisions similar to a human pilot, who mostly flies by the book. Techniques and procedures that it would follow are detailed in public documents. This kind of AI is how ED currently does it, it just needs some improvements.

That's essentially what AI would do.

[MiG21bisFishbedL]

https://en.wikipedia.org/wiki/Computers_are_social_actors

[Dragon1-1]

7 hours ago, Exorcet said:

The brain and AI aren't much different.

You couldn't be more wrong. I'm a biophysicist, so I know a thing or two about brains and computers, and I assure you, they're very different. Let's start with the fact AI runs on a perfectly deterministic machine, so it's 100% deterministic itself. Now, research on "hardware" of the human brain is ongoing, but it's clearly nondeterministic, and probably uses quantum phenomena somewhere, which are encountered in a surprising number of places around a cell. There's just no comparison. AI is really good at fooling us it's thinking, but it really, really isn't. The sooner people realize and accept that, the sooner the BS around AI will stop and serious discussion on how to use it can begin. 

It's not Skynet and never can be, not without a working quantum computer, I suspect. It's simply not possible. A deterministic algorithm has no agency, no imagination, no knowledge (qualities lacking in many humans, too, but I digress). All it does is change some numbers into some other numbers. That's literally all a computer is capable of. We use peripherals to convert those numbers into sounds and images, but under the hood it's still numbers. Humans have a tendency to anthropomorphize, which had caused much dumber systems to seem human (see ELIZA). The only reason it's so good at fooling us is that most human interaction is very superficial and shallow.

7 hours ago, Exorcet said:

I disagree, coding is well within the reach for AI. It's going to take time to develop AI that is good at it, but it's not a far fetched idea.

This will be a separate project, probably expensive as gold (think AlphaFold, not ChatGPT), and in the end, will still require a human programmer, because coding is ultimately creative work. If such an AI is developed, it will be just another programming language, a very, very high level one, but still. Even if it's made within 10 years, it's unlikely to be any better than existing languages in practice, and the only thing it'll be able to do for programmers is acting as a somewhat smarter autocorrect.

7 hours ago, Exorcet said:

That's essentially what AI would do.

Nope, this shows you're not really familiar with how AI works. With AI, you put in a blank slate, and then just put it into situations on which it learns. It won't necessarily replicate human-made procedures, though it might come up with something similar due to convergent evolution. The current "game AI" just follows procedures, and if those procedures are good enough, it can be really good at that. Machine learning has very little to add. This was a discussion a few years back when they made an ML algorithm that could beat a real pilot, and one that beat pros at Starcraft. Nothing had really changed since then. Quite frankly, those algorithms would be a much better fit than anything derived from ChatGPT, but those, too, can't have "preconceptions" put in, so if you want DCS AI to simulate how real pilots fly or flew in historical settings, those models are of no use.

What it could do is use voice synthesis and an LLM to provide more natural comms, combined with voice recognition it could allow casual conversations with AI over radio. However, you'd need to severely constrain the model as to get to sound like a real human in a given role (ATC controller, AWACS, fighter pilot, and so on). That is possibly doable, and would be quite immersive, although it's still a good bit away.

[TheFreshPrince]

vor 2 Stunden schrieb Dragon1-1:

What it could do is use voice synthesis and an LLM to provide more natural comms, combined with voice recognition it could allow casual conversations with AI over radio. However, you'd need to severely constrain the model as to get to sound like a real human in a given role (ATC controller, AWACS, fighter pilot, and so on). That is possibly doable, and would be quite immersive, although it's still a good bit away.

That would be a big step though and really great to have. What I've learned from this thread and further research so far is, that the hype is indeed a bit strong at the moment. It will probably take a bit longer to develop really good and really helpful AI. Things seem to slow down a bit at the moment, likely due to ethical & legal reasons and imminent regulations.

But It's not necessary for AI to exactly function like a human brain to be helpful. All you need to develop is really good machine learning, so that you can train the program to learn your code or whatever you need and get to at least a human skill level. Humans are far from perfect, so also no need to create a perfect AI. Just with a similar error rate or in the best case with a better rate. Then you can easily give the AI tasks and speed up things. Corrections will still be necessary, but right now DCS is buggy as well with human programming.

PS: But I'm glad that ED chose to open up some positions on AI & machine learning. Anything that improves DCS is great!

[Exorcet]

5 hours ago, Dragon1-1 said:

You couldn't be more wrong. I'm a biophysicist, so I know a thing or two about brains and computers, and I assure you, they're very different. Let's start with the fact AI runs on a perfectly deterministic machine, so it's 100% deterministic itself. Now, research on "hardware" of the human brain is ongoing, but it's clearly nondeterministic, and probably uses quantum phenomena somewhere, which are encountered in a surprising number of places around a cell. There's just no comparison. AI is really good at fooling us it's thinking, but it really, really isn't. The sooner people realize and accept that, the sooner the BS around AI will stop and serious discussion on how to use it can begin. 

The physics behind how they work isn't what I'm talking about, more the high level methodology. Both AI and the human brain try to look for patterns. They will take in information and then interpolate or extrapolate from there. This is the idea behind using AI as a light form of physics solver:

Much like a human does, AI will notice trends from "studying" what a fluid will do in certain situations and then be able to make predictions that are generally more accurate with more data. It's not a brain, but the process and end result are similar in a way. Getting back to DCS, AI as a mini CFD solver could be useful in supplementing CFD simulations for reduced cost or as a stand in for CFD all together when total accuracy is less important (AI flight model, flow around buildings, or terrain, animations for trees or other objects in wind, etc).

AI has no agency or consciousness, but that's not important. Just like current AI (as in games and such) has no agency or consciousness but can act like it does if programmed correctly (NPC's in a RPG, etc).

5 hours ago, Dragon1-1 said:

It's not Skynet and never can be, not without a working quantum computer, I suspect. It's simply not possible. A deterministic algorithm has no agency, no imagination, no knowledge (qualities lacking in many humans, too, but I digress). All it does is change some numbers into some other numbers. That's literally all a computer is capable of.

That's all we need. Give the machine data in a format that it can understand and it can start to "see" patterns and then learn to output data that fits those patterns. If you want to consider AI more of a really advanced algorithm rather than intelligence that's fine, but it doesn't change things. Just knowing how to make for loops by itself is something I'd say is pretty helpful. I'd much rather have a single button to press to make one, or make 80% of one than having to code one myself all the time, just for example.

5 hours ago, Dragon1-1 said:

This will be a separate project

Yes it will. Developers aren't going to use ChatGPT, just like AI development going on in my own work will not. It's analogous to the rise of computers where different machines and software will be developed on a per need basis.

5 hours ago, Dragon1-1 said:

, probably expensive as gold (think AlphaFold, not ChatGPT), and in the end, will still require a human programmer, because coding is ultimately creative work.

I won't make estimates on pricing, but the AI work I'm closest to doesn't even seem to be the most expensive part of the equation. In the long term it will likely end up being a cost saver.

As far as not replacing humans, that's totally fine. AI is a tool like any other, and tools don't do much without a user. The AI's job isn't creativity, it's the elimination of repetitive and mundane tasks.

5 hours ago, Dragon1-1 said:

If such an AI is developed, it will be just another programming language, a very, very high level one, but still. Even if it's made within 10 years, it's unlikely to be any better than existing languages in practice, and the only thing it'll be able to do for programmers is acting as a somewhat smarter autocorrect.

I'd take a smarter autocorrect. Predicting the future is difficult, but I'm fairly confident AI is going to fill these roles almost everywhere in the near future.

5 hours ago, Dragon1-1 said:

Nope, this shows you're not really familiar with how AI works. With AI, you put in a blank slate, and then just put it into situations on which it learns. It won't necessarily replicate human-made procedures, though it might come up with something similar due to convergent evolution.

It depends. If you only give it human procedures to learn, that is what it will learn. That's basically what AI language models are doing. They're learning how people talk and then blindly imitating that. The AI doesn't say "good morning" because it wants you to have a nice day, it says it because those words are a common greeting in the inputs it receives.

5 hours ago, Dragon1-1 said:

The current "game AI" just follows procedures, and if those procedures are good enough, it can be really good at that. Machine learning has very little to add.

I disagree here. Machine learning can learn, which is already a huge change. Current AI relies on the developer to update. True AI can update itself and "mutate" into different forms. This is interesting for games because both individual players and the player community as a whole typical evolve in terms of ability as time goes on. It would be nice to have an AI that can progress with players and the overall player pool.

 

5 hours ago, Dragon1-1 said:

This was a discussion a few years back when they made an ML algorithm that could beat a real pilot, and one that beat pros at Starcraft. Nothing had really changed since then. Quite frankly, those algorithms would be a much better fit than anything derived from ChatGPT, but those, too, can't have "preconceptions" put in, so if you want DCS AI to simulate how real pilots fly or flew in historical settings, those models are of no use.

ChatGPT isn't some wondertool. It's specifically designed to model language. The algorithms your mention are closer to what a game developer might use.

5 hours ago, Dragon1-1 said:

What it could do is use voice synthesis and an LLM to provide more natural comms, combined with voice recognition it could allow casual conversations with AI over radio. However, you'd need to severely constrain the model as to get to sound like a real human in a given role (ATC controller, AWACS, fighter pilot, and so on). That is possibly doable, and would be quite immersive, although it's still a good bit away.

Indeed, voices are a huge area of potential for AI. Heatblur apparently lost or misplaced the original files for Jester (basically it's not easy for them to update or add lines I think), but an AI could easily replicate Jester's voice from the final sound files in game and then use the replicate voice to create new lines entirely. Then the AI could even be used to dynamically respond to situations realistically with the new speech.

Something similar can be done with graphics, as AI could study condensation effects or bomb explosions and come up with realistic approximations of how those would look without input from artists outside of the initial learning.

AI in the dynamic campaign or just built into the general AI would also be very interesting. It's one of the ways that ground forces can learn to do more than sit out in the open without the need for programmers to come up with very clever and potentially complicated scripts for determining unit reactions.

[Dragon1-1]

10 hours ago, Exorcet said:

It depends. If you only give it human procedures to learn, that is what it will learn. That's basically what AI language models are doing. They're learning how people talk and then blindly imitating that. The AI doesn't say "good morning" because it wants you to have a nice day, it says it because those words are a common greeting in the inputs it receives.

My issue with that is that you're proposing is hunting mosquitos with a shotgun. It's simply overkill (as much as I'd like to take a shotgun to one of those annoying little bastards). You don't need a learning AI to make an aircraft follow procedures, you need a fancy autopilot. It's far more cost-effective to make an autopilot than an AI, and that's what ED should focus on. There's a book on everything in aviation, and a procedure is essentially an algorithm for humans. What you want is to implement those algorithms with just enough pseudorandom deviation that they don't seem totally robotic. That's how ED currently does its AI. 

Likewise, AI as a CFD solver. Why? You don't need machine learning to do that, you need an efficient algorithm for solving differential equations numerically. Those exist, and there's an entire branch of computer science devoted to developing such algorithms. Yes, an AI can contribute to that, but it's not worth doing that in house.

Remember, AI takes enormous computing resources to train, even if you do have the data. This isn't going to be happening on the players' computers, performance is bad enough as it is. So it wouldn't be learning as the players fight it, this would likely impose an unacceptable performance penalty. ED would have to train it in house and give us a canned algorithm that results. Learning in real time is unlikely to be doable on consumer hardware. 

When applications of AI come up, the first question should always be "can we do this without AI?". Because if the answer to that is yes, then there's usually a much more reliable, less resource intensive path to doing whatever you want to do. There are a precious few problems where the answer is either "no" or "not without a human in the loop". Admittedly, those problems can be big ones, like protein folding, but does this really apply to DCS? 

[Exorcet]

1 hour ago, Dragon1-1 said:

My issue with that is that you're proposing is hunting mosquitos with a shotgun. It's simply overkill (as much as I'd like to take a shotgun to one of those annoying little bastards). You don't need a learning AI to make an aircraft follow procedures, you need a fancy autopilot. It's far more cost-effective to make an autopilot than an AI, and that's what ED should focus on. There's a book on everything in aviation, and a procedure is essentially an algorithm for humans. What you want is to implement those algorithms with just enough pseudorandom deviation that they don't seem totally robotic. That's how ED currently does its AI. 

You're right, AI is not needed to create convincing CPU pilots, but there are potential benefits in involving machine learning in the process. What I don't know is when the crossover point will come. We're not there yet. But as AI becomes better and more accessible, it's going to become a better and better option. I don't expect ED or any other developer to announce ML as a part of their products any time soon, but I would expect them to start looking into how they could incorporate AI into their product development at this point, even at a very shallow level.

 

1 hour ago, Dragon1-1 said:

Likewise, AI as a CFD solver. Why? You don't need machine learning to do that, you need an efficient algorithm for solving differential equations numerically. Those exist, and there's an entire branch of computer science devoted to developing such algorithms. Yes, an AI can contribute to that, but it's not worth doing that in house.

Computers have come a long way, but solving those equations remains costly. AI is there to bring the cost down, and indications are that it is well worth using this method if speed is the concern over total accuracy. That's the gist of the video I posted before and it's the reason why AI is making its way into my own work. Also just to be clear here ED isn't going to be making their own CFD code, AI assisted or not. That's going to be an external product that they buy from someone else.

1 hour ago, Dragon1-1 said:

Remember, AI takes enormous computing resources to train, even if you do have the data. This isn't going to be happening on the players' computers, performance is bad enough as it is. So it wouldn't be learning as the players fight it, this would likely impose an unacceptable performance penalty. ED would have to train it in house and give us a canned algorithm that results. Learning in real time is unlikely to be doable on consumer hardware. 

There are solutions here. The AI can train when the player isn't flying and then use what it learned in the next session. Or the training could indeed be done off of the player machine entirely with the ML being facilitated by sending player data to ED to use and the results sent back to the player.

 

1 hour ago, Dragon1-1 said:

When applications of AI come up, the first question should always be "can we do this without AI?". Because if the answer to that is yes, then there's usually a much more reliable, less resource intensive path to doing whatever you want to do. There are a precious few problems where the answer is either "no" or "not without a human in the loop". Admittedly, those problems can be big ones, like protein folding, but does this really apply to DCS? 

I'm not sure if we should be looking to solve problems without AI. I know that AI shouldn't be looked at as a wonder tool that can do anything, but at the same time I think it's much to early to categorize it in the niche solution bin. Go back far enough and similar discussions were had about computers. They are now practically everywhere because of the one thing they do very well, math. They can do simple calculations at inhuman speeds and their current performance is almost uncountably better than the first models.

I feel like we're at the stage where we don't know what AI will be capable of, so I think it's good to experiment with it. Removing humans from the loop is not the goal, and really doesn't make sense no matter how advanced technology gets because the end goal of technology is to serve human wants. As far as DCS goes, automation is always going to be useful. ED can only hire so many workers and human workers can only work so many hours a day. Even if AI only breaks even with humans on a cost per hour basis and only breaks even on output quality basis, it can at least work continually 24 hours a day. At that point it becomes worth having. It's likely going to be valuable long before then though just by increasing the output of a human worker. Again, I'm not saying this is something ED can snap their fingers and set up right now. They're going to have to research and invest in the technology and then work to get results, but all indications to me are that in the long run it should pay off.

[Dragon1-1]

5 hours ago, Exorcet said:

There are solutions here. The AI can train when the player isn't flying and then use what it learned in the next session. 

Not feasible without DCS running all the time, which I expect people wouldn't want. Also, it'd be inconsistent from player to player, making debugging difficult. Another problem with AI is unrealistic inputs will produce an unrealistic response. Say, a module's FM is bugged, and ED takes its sweet time fixing it, as they often do. AI learns to fight using the broken behavior, against other aircraft using the broken behavior. Then, it suddenly gets fixed. How do you handle that? Do you revert (and lose potentially years of training), or hope it'll eventually unlearn the behavior and withstand a flurry of bug reports about AI being "broken"? Finally, a purely algorithmic solution can't be tweaked to simulate particular historic ways of flying. Difficulty also can't be easily adjusted.

An evolving AI in an evolving environment such as DCS is a bad idea. It'd be constantly playing catch-up (no way to make it read changelogs and adjust on its own) and will spend most of the time "broken". People would be complaining and devs would be powerless to fix it, because AI algorithms are black boxes, a result of training data that can't be adjusted manually. Compared to an algorithm, it's inferior in every way, except maybe the actual ability to win fights using fully mature modules.

5 hours ago, Exorcet said:

I'm not sure if we should be looking to solve problems without AI.

We should, just like we are asking "can this be done without a general purpose computer?" whenever designing a new appliance. Most often, the answer is "yes", and even when it's not, it's mostly because there are already many non-computerized versions on the market. We classify microcontrollers in cars and dishwashers and so on as "computers", but they're only distant relatives to your home PC. Unless you're making a "smart" appliance with the ability to connect to the internet, what you really need is a few simple logic and memory circuits. Using, say, a Raspberry Pi running Linux to drive a washing machine would be overkill, and these things are only done because this kind of tech is ridiculously cheap by now (and even then, a smart washing machine will certainly cost more). It still introduces unnecessary complexity, so most smart devices use a microcontroller much less elaborate than that.

Same with AI, it's usually the least efficient, most difficult to work with approach to the problem. When it comes to problems where it is the only approach, it works great, but for many things, a simple algorithm will likely perform better, be easier to implement and to debug. AI is, in it current state, a niche tool (albeit with some rather important niches), and it's unlikely to become that much more in near term, if only because machine learning had been around for some time, and so the initial rapid development phase has, consequently, already been running for a while. As awesome as the recent advances are, we might be closer to the plateau phase than you think.