Support NeoGAF

[Draugoth]

There are multiple reasons.

Probably the biggest one is that the tech is overhyped and fails to reach developer standards. It's mostly an interesting toy right now that can speed up some a niche of processes.

 

[Sethbacca]

Do you think that has never been tried before?
It just doesnt work!

Something not having worked previously doesn't mean it will never work.

I suppose the other side of this though is that if the AI got too good people would finish even fewer games. Bad AI is good for game design to a degree. You need a certain level of predictability for scenario accounting.

 

[yurinka]

The current GPT models and image and video creation tools are so prolific and are leading to controversies in game development as people use these tools as "shortcuts" while cutting out human writers and artists. People in the gaming community have been using "AI" as a term for decades, but I haven't seen any drastic improvements in "enemy AI" or "companion AI" unless it's just slowly progressing quietly behind the scenes. Honestly, the last time I remember being impressed with AI was MGS2 where guards would notice wet footprints or whatever, or RE4 where the Ganados would dodge your shots or work together to take you out.

I'm not following the industry as close as I used to, so I certainly could have missed something. Lemme know.

TLDR: GPT AI is creating slop, why haven't we seen improved enemy AI or companion AI in games?

We've already seen some working examples, like Darth Vader NPC in Fortnite or a Skyrim mod where you can talk to NPCs and they reply without without the specific 2/3 prefefined quotes.

And not sure who, but I think it was Nvidia, CD Projekt or Ubisoft shown a prototype of the same, not sure if I remember well but including lip synch.

Companies like EA or Ubisoft mentioned to be working using AI to enhance animations, and I think Sony mentioned to use it to make lip synch to all languages with voice over in games and anime.

Some also mentioned to use it to enhance/remaster game textures or models. Other ones like Haven also mentioned to use it to accelerate the creation of in-game wearable customization items.

Basically everybody is trying to see how they can use AI to help them be more efficient and productive in future games. But things, particularly big innovations and changes like this one, take time first to spot the possibilities, second to build the tech and third to make the games using it.

 

[Knightime_X]

Depends on how poorly you want your game to run.
Ai anything isn't free computing.

 

[BlueManifest]

If enemy ai gets too good in games even the best players won't be able to kill them

 

[T4keD0wN]

For example in a game like GTA, a smart AI would just leave your Wanted rating at 5 stars all the time and not stop until you are dead.

So a state that doesnt end (to move into another state) or adjusts when conditions are met? Thats the exact opposite of smart, its less advanced, simpler than the original.

Ill do you one better, an endless loop of game over screens when you boot up the game, really hard to beat, quite possibly the smartest AI ever (if we use your defition of AI) could fit on a floppy.

But then it isnt smart, as a smart AI wouldnt want to get defeated and pop their head out. The AI is kept artificially dumb to keep it fun for the player.

How can it defeat the player then, if it need to not have a condition (being in cover here) to perform a function (attack player)? With no input from the player this would make them game functionally stuck. You could put the gamepad down for 3 days, come back and your example of "smart AI" literally wouldnt be able to win.

No I´m not, ive been designing games professionally for 20+ years and worked with various AI systems like AI Implant.
There simply isnt any "smart" AI in any games, it is always heavily scripted and is mostly smoke and mirrors.

A game designer with 20 years of experience who appears to think that AI in games thinks, believes that there is supposed to be a difference between a collection of scripts and AI, and doesnt know a difference between having functions/scripts/systems and difficulty tuning?

Sounds unbelievable, among other giveaways, why dig the hole for yourself? Had what you claim been true you wouldnt say this:

mixing up good game design and smart AI

because youd actually know what youre saying there given that "AI" is designed.

The problem was that players played too sloppy and the AI noticed those mistakes.
So ultimately they had to make the AI dumb and overlook the players mistakes to keep it fun!

Oh really, AI noticed how they play and evaluated things as mistakes? Weird, i didnt know its possible for a script to "know" whats a mistake. The closest it can do is read values to call functions based on conditions.

 

[Miyazaki’s Slave]

If you do not understand how the tools are used you are unable to see the improvements.

 

[Sakura]

But then it isnt smart, as a smart AI wouldnt want to get defeated and pop their head out. The AI is kept artificially dumb to keep it fun for the player.
Like take chess for example - is it fun to play against the much smarter AI? Or is it more fun to play against a human who is on your level?
If you know you cant win, would you even want to play?

This is kind of a load of shit. People always bring up arguments like this when there is discussion about better AI in games.
People want AI that is more believable, or in other words, acts more like a human.
If you are playing a PVP game, lets say Halo mp or Counter Strike or something, do humans just hide behind cover and then never pop out because they don't want to die?
Even in the real world people don't act like that. They have orders to follow, goals to achieve, etc.
If your AI enemy is just sitting behind cover forever, then it has been poorly designed.
It should be using visual and audio queues to determine when it should make a move. Or it should coordinate with the other enemy units. Or it could try throwing a grenade or something to try to get you out of your cover, then pop it's head as it expects you to make a run for it.
If your "smart" AI was just sitting behind cover refusing to peak, then you could just walk up to it hiding and kill it. Not very smart huh?

Please keep in mind that no matter which game you're playing, the "AI" knows where you are and can read all your moves at the input level so it is by default 100000000 times better than you.

the real struggle is dumbing down the AI to a level that it feels challenging and rewarding to beat it, without it either being too brutal or too lenient. Proper AI programming is brutally difficult which is why so many games fail at it.

The AI does not need to know where you are and read all your inputs, that's not how AI works lol.
You can do even basic shit like implement vision cones for the AI, and have it only have knowledge of what it "sees". The AI doesn't know your location or what inputs you are making, without even seeing you, unless it has been programmed to.

 

[realcool]

Where Winds Meet has AI NPC's and it can be hilarious when you know how to manipulate them to get what you want

Me:

 

[ScHlAuChi]

So a state that doesnt end (to move into another state) or adjusts when conditions are met? Thats the exact opposite of smart, its less advanced, simpler than the original.

The GTA example was used to say that what a human considers a smart action is not compatible with video game AI logic.

Ill do you one better, an endless loop of game over screens when you boot up the game, really hard to beat, quite possibly the smartest AI ever (if we use your defition of AI) could fit on a floppy.

I have no idea what this example has to do with AI...

How can it defeat the player then, if it need to not have a condition (being in cover here) to perform a function (attack player)? With no input from the player this would make them game functionally stuck. You could put the gamepad down for 3 days, come back and your example of "smart AI" literally wouldnt be able to win.

You still dont get it - I was not talking about how the AI is scripted, I was talking about how a human would judge the smartness of the AI.
If the AI was a thinking self acting smart actor, it would stay in cover to not die. But then you wouldnt have a game!
But the AI is dumb and stands up to get shot! That is not smart!

A game designer with 20 years of experience who appears to think that AI in games thinks, believes that there is supposed to be a difference between a collection of scripts and AI, and doesnt know a difference between having functions/scripts/systems and difficulty tuning?

Huh did you not read what I wrote? I said the exact opposite, that AI in games is NEVER smart!
The OP is basically asking for smart AI that can "think" to appear smart, and I have been arguing that this wont ever happen!
Even if you have an AI with hundreds of scripts that react to certain things, it doesnt mean its actually smart in the sense the OP expects!
And none of this has anything to do with difficulty!
Also I dont give a flying fuck if you question my experience, I know what I worked on, and you wont change that

Sounds unbelievable, among other giveaways, why dig the hole for yourself? Had what you claim been true you wouldnt say this:
because youd actually know what youre saying there given that "AI" is designed.
Oh really, AI noticed how they play and evaluated things as mistakes? Weird, i didnt know its possible for a script to "know" whats a mistake. The closest it can do is read values to call functions based on conditions.

That is exactly what I said, AI is designed and scripted and is never smart!
And you once again misinterpreted this, I did not say the AI interpreted how they play, the AI was scripted to individually act per enemy character to many stimuli.
But already that was too annoying to players, so a real actual smart AI would be even less fun!
Im not going to continue wasting my time discussing with you when you argue in bad faith.

 

[Danny Dudekisser]

"AI" as people are using it these days isn't really the same as the AI that controls enemy behavior. As some others have pointed out, AI isn't really something where you want the enemies to be as smart or capable as a real player. That just wouldn't be fun. Enemy AI improvements will come from developers consciously deciding to be deliberate about setting routines, creating level designs where they can come up with novel routines that the player won't expect (like FEAR's guards working together to flank you), and putting in a lot of effort to constantly build new surprises. It's as much a part of game design as building the level layouts or the mechanics.

But it's hard work and takes a lot of effort. Guess what modern dev teams, especially in AAA, don't like.

 

[yet_another_alt]

I saw some of those Where Winds Meets convos. They're hilarious.

 

[MrAlwaysRight]

This is interesting, thanks for the clarification between the LLM and the coded "AI", but couldn't you "train" bots or enemies to react based on player inputs using that same idea? Instead of language it's player inputs or location data within the game. I'm not the most technical person, so the explanations are helpful.

Well an LLM is prediction, it's very sophisticated, and the training is so deep that we don't really understand how it works, we understand the mechanics of the transformer architecture, the attention mechanisms and how tokens flow through layers from a mathematical perspective, but we don't really understand how that resolves in logic and reasoning, from a design point of view it's very sophisticated mathematics that resolves in something that appears to be artificial intelligence.

Training a game's bot to work on the same type of prediction but based on an input is possible, but it would basically be cheating because it would respond based on what direction and button you pressed, it's a much smaller pool of inputs compared to an LLM so training would be simple, but it wouldn't be fun to play against.

Where we could use LLM's is for NPC dialogue in games, but consoles are a way off that yet, we have some tiny models that could realistically run on a PS5 Pro but it would probably still be quite slow and irksome.

 

[Pasta la Vista]

Arc Raiders has some very interesting AI enemies.

 

[existensmaximum]

If AI is so great, why don't you just ask it?

 

[genesisblock]

Here´s a little secret - gamers do NOT want smart AI!
Gamers want to win and have power fantasies - smart AI makes the game annoying and hard.
No one wants to play that!

Uhh? There's only one problem with your statement. There is no game out there with smart AI since Killzone 2, Halo and F.E.A.R. And it only has enhanced the experiences not diminished it. Give us examples of these "supposed" frustrating-but-smart-AI and not just a "buffed" NPC.

 

[Radical_3d]

I mean you can get around some of this by putting in guardrails around the questions, query/prompt caching, etc… but yeah, you are going to be making a cloud call to a data center as a complex AI isn't going to run on even a PS6.

Yeah, conversation wise you could work around it, but OP is talking about enemy AI in the traditional sense. Path finding, combat strategies and such. That can't be done in your average German ADSL crappy connection, worse than that, those of us who actually have good connections still are not there in terms of latency. That must be thought in milliseconds. What the OP is asking for is no other thing than the power of the cloud. An Xbox One feature! And we all saw how good that went in Crackdown 3 (it was finally not implemented for those who don't remember).

The solution is having better thought games with more budget of better CPUs for your games. So basically your enemies are: the aging CPUs of the past two generation of consoles and, even more crucially, this:

Every gain we have in every generation of CPUs is swallowed by gains in graphical fidelity. And then, of course, is the design choice. Would you make your enemies smarter and your game harder as a baseline (since it's the AI, that is not easily scalable) when the trophy data shows you that only 5% of gamers had finished your previous game?

 

[Negotiator101]

Because what people call "good AI" in games is different from the kind of AI these models can produce.

You don't want adaptativeness and skill on the NPCs, you want reactivity. This latter one still relies almost entirely on the developer's foresight.

I'm sure reading many years ago without scripting it's incredibly hard to get a AI character to do even very simple things, example picking up a object. Is that still true today?

 

[icecold1983]

Because most gamers don't actually want good AI. Good AI would be the same as playing online. You all get wrecked online and start whining about "try hards."

 

[StereoVsn]

Yeah, conversation wise you could work around it, but OP is talking about enemy AI in the traditional sense. Path finding, combat strategies and such. That can't be done in your average German ADSL crappy connection, worse than that, those of us who actually have good connections still are not there in terms of latency. That must be thought in milliseconds. What the OP is asking for is no other thing than the power of the cloud. An Xbox One feature! And we all saw how good that went in Crackdown 3 (it was finally not implemented for those who don't remember).

The solution is having better thoughts games with more budget of better CPUs for your games. So basically your enemies are: the aging CPUs of the past two generation of consoles and, even more crucially, this:

Every gain we have in every generation of CPUs is swallowed by gains in graphical fidelity. And then, of course, is the design choice. Would you make your enemies smarter and your game harder as a baseline (since it's the AI, that is not easily scalable) when the trophy data shows you that only 5% of gamers had finished your previous game?

Oh, good point. I was mainly thinking conversational AI, NPC scheduling, those sorts of things.

For enemy AI behavior, pathing, etc… even difficulty aside, does not seem vast majority of devs are interested. We have had better enemy AI in a lot of older games.

Same goes for physics based environmental destruction. It is all unfortunate as devs could make things more interesting.

 

[Sentenza]

The short answer is: "because no one is attempting to use AI to make games better. They are just trying to use it to make the production process cheaper".
And the quality of the result reflects the intent.

 

[ScHlAuChi]

Uhh? There's only one problem with your statement. There is no game out there with smart AI since Killzone 2, Halo and F.E.A.R. And it only has enhanced the experiences not diminished it. Give us examples of these "supposed" frustrating-but-smart-AI and not just a "buffed" NPC.

There is no problem with the statement, as there is no game out there that has smart AI!
None of those games you mentioned had it either, thats just a combination of careful AI scripting and good leveldesign.
Add in convincing "search" animations and the player thinks the AI is smart while in reality it is all just smoke and mirrors.

 

[RoboFu]

I disagree when people say ai is getting better. Beyond making dumb images and videos I have yet to see ai be really useful in the workplace beyond a search engine. I've tried to use ai for coding and at best it can be used as a refresher to jog your memory. Straight up copy and paste either never works or is the worst possible way to do something.
AI isn't really thinking .. it's just picking things from the internet that's the most popular. Youn can 9 times out 10 find the stack exchange post that ai is pulling code directly from .

 

[Poppinfresh]

The autistic screeching that happens every time a developer tries to use AI creative assets or AI Voice-overs in a game is probably a pretty big impediment to developers seriously considering widespread AI use of any sort in their games.

 

[Guilty_AI]

I'm sure reading many years ago without scripting it's incredibly hard to get a AI character to do even very simple things, example picking up a object. Is that still true today?

If by scripting (thats an incredibly vague term for game dev) you mean making character animations + a series of logical steps do do a specific action, then yeah, pretty much still the same.

 

[Fabieter]

People underestimate that really really good ai wouldnt be fun to play against.

Imagine a from software boss which learns from all of your cheesy tactics.

People would cry that its not fun

 

[Minsc]

AI peaked in 1993 when 7th Guest released. Games still haven't topped that one.

 

[Shtof]

I think the next 10 years will see an explosion of improving game opponent/enemy AI.
But it will not come in the form of LLMs or traditional video game AI techniques.
Rather, it will be from agentic teammates or assistants that will follow you between games and world generators being able design entire gameplay paradigms on the fly.
I'll be rather surprised if PS6 doesn't have this as a launch feature.

 

[Heimdall_Xtreme]

In Perfect Dark, the AI is so brutal that you could be killed with a single, accurate shot from a distance.

 

[Humdinger]

I'm guessing one reason is that the tech is just recently developing and games take many years to make. We won't see it right away.

 

[Nintendo's Witch]

LLMs are not artificial intelligence

 

[Topher]

LLMs are not artificial intelligence

It is an AI system or subset.

 

[Topher]

I disagree when people say ai is getting better. Beyond making dumb images and videos I have yet to see ai be really useful in the workplace beyond a search engine. I've tried to use ai for coding and at best it can be used as a refresher to jog your memory. Straight up copy and paste either never works or is the worst possible way to do something.
AI isn't really thinking .. it's just picking things from the internet that's the most popular. Youn can 9 times out 10 find the stack exchange post that ai is pulling code directly from .

I personally get answers to my coding problems/questions a hell of a lot faster with ChatGPT then I do searching dozens of websites. I could take a screenshot of this reply control in GAF and copy it into Copilot and have it generate it in HTML.

Actually, I'll do that.

Here is the screenshot:

This is a screenshot of the HTML that Copilot created based on that image:

That's pretty dang useful. And that isn't even scratching the surface of really long mundane tasks (like analyzing log files and such) and other things humans are not very good at in IT and other fields.

 

[RoboFu]

I personally get answers to my coding problems/questions a hell of a lot faster with ChatGPT then I do searching dozens of websites. I could take a screenshot of this reply control in GAF and copy it into Copilot and have it generate it in HTML.

Actually, I'll do that.

Here is the screenshot:

This is a screenshot of the HTML that Copilot created based on that image:

That's pretty dang useful. And that isn't even scratching the surface of really long mundane tasks (like analyzing log files and such) and other things humans are not very good at in IT and other fields.

Isn't that what I said ? It's basically a better search engine. It will do that stack exchange search for you but it almost always brings up THE ONE MOST POPULAR POST of what you are searching for. That can be useful but again it's almost never a full answer or best practices.

 

[Topher]

Isn't that what I said ? It's basically a better search engine. It will do that stack exchange search for you but it almost always brings up THE ONE MOST POPULAR POST of what you are searching for. That can be useful but again it's almost never a full answer or best practices.

No, it isn't. AI can run entirely offline. I have AI running in LM Studio on my linux box. I gave it a very specific code task and it cranked out the code.

 

[Sentenza]

People underestimate that really really good ai wouldnt be fun to play against.

"Good AI" in videogames terms is AI that behave convincingly for what it is trying to be.
It's not AI that competes with the player.
Otherwise unbeatable insta-kill AIMBOTs would have been a thing 20 years ago already.

 

[RoboFu]

No, it isn't. AI can run entirely offline. I have AI running in LM Studio on my linux box. I gave it a very specific code task and it cranked out the code.

Ehh that's a very basic operation though using the standard libs. I mean my 10 year son can weight a for each loop with set list.

Now do it .. even in c# .. using a third party frame . Even a popular one like using GTK with something a little more complexed like implementing a flood fill on an indexed color bitmap that will work on windows and Linux.

Even with that you will get SOMETHING but I doubt it works without some tweaking and it will probably be the slowest and worst possible way possible.

 

[Topher]

Ehh that's a very basic operation though using the standard libs. I mean my 10 year son can weight a for each loop with set list.

Now do it .. even in c# .. using a third party frame . Even a popular one like using GTK with something a little more complexed like implementing a flood fill on an indexed color bitmap that will work on windows and Linux.

Even with that you will get SOMETHING but I doubt it works without some tweaking and it will probably be the slowest and worst possible way possible.

The complexity of the program isn't the point. This was showing you that it isn't just finding the most popular post on the internet. This was done offline so it isn't just a "better search engine".

Obviously there are more complex tasks than this that it is capable of. I have plenty of use cases at work, but doubt it would make much difference at this point.

 

[genesisblock]

There is no problem with the statement, as there is no game out there that has smart AI!
None of those games you mentioned had it either, thats just a combination of careful AI scripting and good leveldesign.
Add in convincing "search" animations and the player thinks the AI is smart while in reality it is all just smoke and mirrors.

That's what "smart" AI is, and the direction developers should be advancing instead of regressing. It gives the impression of intelligence while providing a challenge to the player. Simply downplaying their achievements to your standard enemy AI grunts only highlights the flaw of your argument. The onus is on you to provide clear examples where smart AI frustrates the gamer in any kind of context before arriving into that conclusion.

 

[dgrdsv]

Cause gamers don't want "better AI". What they want is a "fun AI" and NN based AI isn't there yet to be more fun than what games have now.

 

[Felessan]

When you see how people actively using "cheats" (prompting AI for it to persuade itself) to bypass LLM chatting with NPC for meager little bonuses - it's very obvious just how much general crowd embrace this idea of "better AI"

Happily nothing serious (afaik) is behind this stuff so I just can skip it.

 

[neocycle]

It will come when most people have devices that allow it. I predict next gen will be all about AI.

 

[grvg]

Because no one wants actual AI as enemies in games. You would just always lose unless the AI is programmed to play dumb, and then it's no different than what we have now. And in fcat even now, in many games, you only win because the AI is programmed to let you win if you perform at a certain level. Take fps games for example. A computer does not miss, ever, every time an enemy misses you in an fps game it's because it was programmed to do it. It could just as well be programmed to hit you every aingle time the very milisecond you get out of cover.

See: the cultists from Blood

 

[ScHlAuChi]

That's what "smart" AI is, and the direction developers should be advancing instead of regressing. It gives the impression of intelligence while providing a challenge to the player.

That is not the smart AI the OP was talking about.
And giving the impression of intelligence when it actually isnt, is exactly the smoke and mirrors I was talking about.
You assume its more challenging because it appears that way. But it is all in your head!

Simply downplaying their achievements to your standard enemy AI grunts only highlights the flaw of your argument.

There is nothing to downplay there when one knows how those AI´s work. Once you know how the stage magician does his tricks, the magic is gone.

The onus is on you to provide clear examples where smart AI frustrates the gamer in any kind of context before arriving into that conclusion.

How can I give clear examples when none such exist? If good AI was so much fun then where are all the games that have it?
And I dont care if you agree to my conclusions or not, I just gave you my opinion as someone who worked on those things.
You are free to believe whatever you want and come to your own conclusions!

 

[Hudo]

It becomes a bit more complicated when you only have a couple of ms for inference (the rest of the allotted ms per frame are already taken up by other systems, like game logic, renderer, sound, network, etc.). That usually means compromising by reducing the model size (which will reduce the model's capabilities, because we already deal with fairly small models for that purpose). And now imagine you either want to have one model for inference for each NPC or a lot of very small models for each NPC or something. If you choose how you utilize models carefully, you could totally make some cool shit. I hope that's what Valve are doing with HLX. Doesn't necessarily have to be restricted to NPC behavior. Could also be models for physics, terrain/level deformation/generation. Or whatever.

Which is why Yann LeCun et al. currently hype up World Models, which are just big-ass models that generate the whole game. The main limitations there right now are context window and "rendering" resolution/fidelity, afaik.

 

[Autistic_Pancake]

> why aren't we getting better "AI" in games?

"AI" in games and LLMs are entirely different things. Most gamers don't have enough processing power to handle LLMs embedded into games, if that's what you're wishing for.