This is our first foray into discussing how machine learning works here on AI 101. This is long overdue but we're going to be doing it slowly and gradually, given there's a LOT of ground to cover. In the meantime, check out the materials recommended in the description if you want to learn more!
Honestly symbolic AI as a means of getting QA data is super effective. Used to work on a mobile game where we collected instantaneous playtest data (number of wins, losses, moves left, etc.) from an AI bot, then played the level ourselves 10 times. On average we knew the bot would be a little worse at the game than a skilled player (about 2-3 more losses / 10 games) but it was consistent enough to be useful. Basically, it helped us know exactly where our levels fell in terms of difficulty and if they were fair (not swingy and up to chance). Another project I wrote a symbolic AI to playtest a co-op mode of a game I made since it's kinda impossible to do that alone, when you're an indie with no roomates who game.
this right after I watched a SQEX GDC about creating a system with GA to test party composition and find out balance breakers. (of course it was about a mobile gatcha game and deemed "working as intended". But we're talking about technics, not how the industry will use them ethically)
Man, you really need to take a look at the article Generative Agents: Interactive Simulacra of Human Behavior! Great news for the future of human simulate behavior... great video btw.
I started out as a hobbyist game developer and now studying AI in my college. With all the things going on with NLP and Generative AI, I was confused with what I actually want to use my knowledge of AI for. I watched all these videos about what ML, DL and others can be used for real-life practical projects but somehow, that did not interest me that much more than just "well, that's interesting. okay, anyways" I KNEW IT. I like Game Development much more than anything. Watching this video made me realized that my passion resides in Game Development Industry, not in general software development. I like how Animation Blending is used. And Texture Upscaling? You mean I can get low res textures from Asset Store and turnt into high res ones? I DIDN'T REALIZE THAT. More than that, real time texture upscaling DLS3??? Holy cow And QA is like how Selenium is used in browser testing. I want to know how it actually model player behaviour tho. Isn't player behaving like a NPC in that case as well? Anyways, thank you for the video and thanks for making me realized my passion again :D Subscribed!
I just discovered your channel and I am very impressed by the quality of your content. All I watch are videos on game development and machine learning, so I don't know how the algo is just recommending you now. Time for some binge watching! :)
If u mean rimworld man I have 30 hours in that games and the only thing that happened are ppl at king my base, animals atacking my base or my ppl dying randomly lol
@@ALexalex-ss4sb No, they're talking about Rain World. That's why they wrote "Rain World." If they were talking about RimWorld they would have written "RimWorld." That's how words work.
text generated quests could be a thing. Let the AI smash together fantasy tropes to generate quests that a software can use to populate the landscape with stuff.
LLM's future gaming? How about an AI game where you interact by talking to it. You set the scene or type of scenerio you want to be immersed in. The game finds info about you from Social media, etc. Puts you as the main character perhaps with people you actually know. In your environment, streets, places, etc. Or a totally made up or AI created scenario based on the initial scene creation. It may include real life events it finds from the Web, including news feeds, and port these into the game. It may build a profile of you and set the gameplay to suit your character, skills, knowledge, etc. Mixing reality with fantasy. Imagine this through Apples Vision Pro! (Yh, I know, privacy concerns). This can also be used in Training of all sorts, interactive like never before.
Incredible coincidence, I'm an artist researching about game ai for a comic I'm working on. The main characters work on a vr/ar system that does exactly what you described, but as a way of therapy. It basically offers individualized therapy of different kinds by interwaving it into a unique story tailored to the player. It generates backgrounds and characters, and as you play through the story, you learn strategies for coping with your problems and stuff, based on existing therapy methods.
@@thechosenegg9340 the trope of Sphere by Michael Chrichton, or Solaris or Stalker by Tarkovsky. A narrative device pulls out the unconscious mind. In your case, there are years of research or papers published in pubmed about Virtual Reality and PTSD, as I checked about that idea years ago. As seductive might sound a therapy based in virtually presence traumatic elements through desensitization, some branches of therapy could argue that the therapy work would be through the vocalization of the inner symbolism on that matter, as you could do making to talk a LLM, finding out what are its biases, and being the case of the patient the way to move the "trauma" out of repression.
Great content - this is the subject that interests me most, AI machine learning etc, and I'm seeing how it works better in some scenarios than others. MotoGP games have been using ANNA, their neural learning AI for the past few games but as an avid user of the series, I don't see much improvement in many ways, they're worse than the old symbolic type AI. Sony's Sophy in GT7 was much better but, it was tested in a small environment of a few cars on a few tracks and while impressive, I wonder how it would perform across the board with the vast number of different vehicles and more tracks? I think AI machine learning is going to be more forceful, as you have pointed out, in the development of games and gaming systems and concepts, than as opponents, for they will always lack the spontaneity of playing or racing against real human minds, at least as it stands now. We as humans can choose to do the unexpected, the unpredictable and the outrageous - sometimes this works other times, not... but to me, it is our fallibility as well as ur curiosity of "lets see if this works" that makes us human.
Literally just watched a video saying the so in hello neighbor 2 literally works the same as the basic so they had in the alpha so I don't think it belongs on the list tbh
Game developers want control over their creation to reduce the probability of bugs. Especially with open worlds, , they don't want all the NPC to jump around until they all gater in a hole, because it's the more effective way of surviving. What machine learning consider optimal may ruin the fun. I can see a lot of applications for machine learning, like bettering the visuals or doing certain specific tasks, but it needs to be predictable.
"Given the opportunity, players will optimize the fun out of a game." If the most effective way of surviving in your game, is to jump around and gater in a hole, players will find that, and play it like that. Using bots to test a level is just another way to find out places that may need more work done.
@@Inferryu "If the most effective way of surviving in your game, is to jump around and gater in a hole, players will find that, and play it like that." The best way to survive Doom is to stay where your character just spawned and go read a book. Does it mean that you have to redo your whole system? I don't think so...
If that is an issue, then we need to redefine what "success" is to the AI when we're training it. The problem isn't that that is actually optimal, the problem is that we haven't rewarded any other behaviors other than optimizing sheer time of life.
That... is the accuracy? Human overwatch reviewers are still doing the convicting--the stat is just comparing the conviction rates between *_AI-reported_* cheaters and *_human-reported_* cheaters. So human reviewers found that AI-generated reports are 80%-95% accurate, while human reports are 15%-30% accurate.
Ultimately that's both the goal and the flaw. You give it a goal and provided the training environment is satisfactory, it will achieve optimal results. But if the goal is unclear, or the goal changes, then it needs to adapt. And right now ML algorithms don't adjust to change easily.
@@AIandGames Yeah, I think we'll need to change the design to break from the old methods. But that's not something I'm able to do, so I'll wait and see.
The deep learning explanation is not really accurate. Deep learning is simply machine learning where you use deep neural networks, which are neural networks with many layers. It doesn't have to be supervised learning, and most reinforcement learning is not deep learning (although there is a subfield called deep reinforcement learning which is part of deep learning).
Publisher: Finally, we can develop video games without those uppity developers! Okay AI, tell us the magic secrets! AI: [Make games with compelling gameplay for single players first and foremost and no intention of charging the players extra unless you are developing new content] Publisher: *throws computer out window*
That industry is mostly dying and getting more and more anti consumer. So, big editors abusing AI won't get us better games, just more and more cloned content that no one even made.
@@fritt_wastaken I suppose the thinking here is that by leaving a lot of the gamedev process up to automated processes, indie devs would be able to make games with less effort, time and costs?
Looking forward to seeing where this goes, even as the clinate around AI hardens as unethical people figured out how to do it to rip off the skills of artists and claim it as their own, causing a whole anti-AI art backlash, and more and more people coming to see AI in a lot of fields as a job-stealing measure that the wealthy will use to sabotage the workers. In between the many upset voices clamoring for it's downfall and the same tech bro's who once were sooo hype on NFT's now turning their eyes towards AI as their latest grift which they praise endlessly and vapidly, I see this channel as a good way to learn what AI actually DOES in tech... and figure out whether or not I should actually be worried.
I appreciate the support. I feel it's important to stay grounded and be realistic about how this technology works. Right now there's a lot of noise and hype around AI, and a lot of that noise is specifically about AI for games that is, frankly, unfounded and based on naïve understandings of how game development works (the Web3 bros are coming for the AI now). It's frustrating to watch, and I prefer to focus on what actually works, rather than speculation and hype. As with every video I make, I can point to the source and why it was built like that. I have a rant of sorts that is brewing in me, that I may well write up and publish on my second channel, given I think we're heading for a new wave of cash-grab nonsense asset flips that are all rely on AI-generated pipelines. Meanwhile there will continue to be smart and effective use of ML in game productions to achieve real results, that still understands the need for having humans in the loop and more traditional design tools.
This is our first foray into discussing how machine learning works here on AI 101. This is long overdue but we're going to be doing it slowly and gradually, given there's a LOT of ground to cover. In the meantime, check out the materials recommended in the description if you want to learn more!
Thanks!
007
ءشعجارالزخزن مرحبا
Honestly symbolic AI as a means of getting QA data is super effective. Used to work on a mobile game where we collected instantaneous playtest data (number of wins, losses, moves left, etc.) from an AI bot, then played the level ourselves 10 times. On average we knew the bot would be a little worse at the game than a skilled player (about 2-3 more losses / 10 games) but it was consistent enough to be useful. Basically, it helped us know exactly where our levels fell in terms of difficulty and if they were fair (not swingy and up to chance). Another project I wrote a symbolic AI to playtest a co-op mode of a game I made since it's kinda impossible to do that alone, when you're an indie with no roomates who game.
I love how, aside from the Halo change, Hello Neighbor 2 was just randomly added to the thumbnail and I find that extremely funny
It's our little secret. 😉
this right after I watched a SQEX GDC about creating a system with GA to test party composition and find out balance breakers.
(of course it was about a mobile gatcha game and deemed "working as intended". But we're talking about technics, not how the industry will use them ethically)
Man, you really need to take a look at the article Generative Agents: Interactive Simulacra of Human Behavior! Great news for the future of human simulate behavior... great video btw.
I started out as a hobbyist game developer and now studying AI in my college. With all the things going on with NLP and Generative AI, I was confused with what I actually want to use my knowledge of AI for. I watched all these videos about what ML, DL and others can be used for real-life practical projects but somehow, that did not interest me that much more than just "well, that's interesting. okay, anyways"
I KNEW IT. I like Game Development much more than anything. Watching this video made me realized that my passion resides in Game Development Industry, not in general software development.
I like how Animation Blending is used.
And Texture Upscaling? You mean I can get low res textures from Asset Store and turnt into high res ones? I DIDN'T REALIZE THAT. More than that, real time texture upscaling DLS3??? Holy cow
And QA is like how Selenium is used in browser testing. I want to know how it actually model player behaviour tho. Isn't player behaving like a NPC in that case as well?
Anyways, thank you for the video and thanks for making me realized my passion again :D
Subscribed!
I just discovered your channel and I am very impressed by the quality of your content. All I watch are videos on game development and machine learning, so I don't know how the algo is just recommending you now. Time for some binge watching! :)
Can you explain Rain World's AI? I've heard it's very complex and interesting in a sense of how it works.
I think there's already videos on this, it's still cool and a good idea though
If u mean rimworld man I have 30 hours in that games and the only thing that happened are ppl at king my base, animals atacking my base or my ppl dying randomly lol
@@ALexalex-ss4sb No, I have rimworld too, just google rain world. It's cool
@@ALexalex-ss4sb No, they're talking about Rain World. That's why they wrote "Rain World." If they were talking about RimWorld they would have written "RimWorld." That's how words work.
text generated quests could be a thing.
Let the AI smash together fantasy tropes to generate quests that a software can use to populate the landscape with stuff.
I think they must have used that for fallout 4
@@gordonramsdale That's basically radiant quests in FO4 and Skryim, yes.
LLM's future gaming? How about an AI game where you interact by talking to it. You set the scene or type of scenerio you want to be immersed in. The game finds info about you from Social media, etc. Puts you as the main character perhaps with people you actually know. In your environment, streets, places, etc. Or a totally made up or AI created scenario based on the initial scene creation. It may include real life events it finds from the Web, including news feeds, and port these into the game. It may build a profile of you and set the gameplay to suit your character, skills, knowledge, etc. Mixing reality with fantasy. Imagine this through Apples Vision Pro!
(Yh, I know, privacy concerns).
This can also be used in Training of all sorts, interactive like never before.
Incredible coincidence, I'm an artist researching about game ai for a comic I'm working on. The main characters work on a vr/ar system that does exactly what you described, but as a way of therapy. It basically offers individualized therapy of different kinds by interwaving it into a unique story tailored to the player. It generates backgrounds and characters, and as you play through the story, you learn strategies for coping with your problems and stuff, based on existing therapy methods.
@@thechosenegg9340 That sounds like a great idea, using it as a therapy tool is a good use of such tech. Keep us posted on the progress 👍
@@thechosenegg9340 the trope of Sphere by Michael Chrichton, or Solaris or Stalker by Tarkovsky. A narrative device pulls out the unconscious mind. In your case, there are years of research or papers published in pubmed about Virtual Reality and PTSD, as I checked about that idea years ago. As seductive might sound a therapy based in virtually presence traumatic elements through desensitization, some branches of therapy could argue that the therapy work would be through the vocalization of the inner symbolism on that matter, as you could do making to talk a LLM, finding out what are its biases, and being the case of the patient the way to move the "trauma" out of repression.
Great content - this is the subject that interests me most, AI machine learning etc, and I'm seeing how it works better in some scenarios than others. MotoGP games have been using ANNA, their neural learning AI for the past few games but as an avid user of the series, I don't see much improvement in many ways, they're worse than the old symbolic type AI. Sony's Sophy in GT7 was much better but, it was tested in a small environment of a few cars on a few tracks and while impressive, I wonder how it would perform across the board with the vast number of different vehicles and more tracks? I think AI machine learning is going to be more forceful, as you have pointed out, in the development of games and gaming systems and concepts, than as opponents, for they will always lack the spontaneity of playing or racing against real human minds, at least as it stands now. We as humans can choose to do the unexpected, the unpredictable and the outrageous - sometimes this works other times, not... but to me, it is our fallibility as well as ur curiosity of "lets see if this works" that makes us human.
@AIandGames. How do you think Ubisoft could use their Watch Dogs Legion character algorithm for other titles like Assassin's Creed?
Nice to be in the early squad once again
Very educational, thank you
Literally just watched a video saying the so in hello neighbor 2 literally works the same as the basic so they had in the alpha so I don't think it belongs on the list tbh
Awesome video, as always
Game developers want control over their creation to reduce the probability of bugs. Especially with open worlds, , they don't want all the NPC to jump around until they all gater in a hole, because it's the more effective way of surviving. What machine learning consider optimal may ruin the fun.
I can see a lot of applications for machine learning, like bettering the visuals or doing certain specific tasks, but it needs to be predictable.
"Given the opportunity, players will optimize the fun out of a game."
If the most effective way of surviving in your game, is to jump around and gater in a hole, players will find that, and play it like that. Using bots to test a level is just another way to find out places that may need more work done.
@@Inferryu "If the most effective way of surviving in your game, is to jump around and gater in a hole, players will find that, and play it like that."
The best way to survive Doom is to stay where your character just spawned and go read a book. Does it mean that you have to redo your whole system? I don't think so...
If that is an issue, then we need to redefine what "success" is to the AI when we're training it. The problem isn't that that is actually optimal, the problem is that we haven't rewarded any other behaviors other than optimizing sheer time of life.
16:37 I don't understand. You make no mention of accuracy when comparing those two different conviction rates.
That... is the accuracy? Human overwatch reviewers are still doing the convicting--the stat is just comparing the conviction rates between *_AI-reported_* cheaters and *_human-reported_* cheaters. So human reviewers found that AI-generated reports are 80%-95% accurate, while human reports are 15%-30% accurate.
Thank you for your amazing work
Current Gpu cant run a LLM and a big Game at the same time locally.
Not for long....
Thanks !
Chatgpt 4.2 connected with other systems will change a lot in game industry.
ML seems weirdly inflexible after it's set. I personally find it weird how little it would adapt despite the goal of it.
Ultimately that's both the goal and the flaw. You give it a goal and provided the training environment is satisfactory, it will achieve optimal results. But if the goal is unclear, or the goal changes, then it needs to adapt. And right now ML algorithms don't adjust to change easily.
@@AIandGames Yeah, I think we'll need to change the design to break from the old methods. But that's not something I'm able to do, so I'll wait and see.
The deep learning explanation is not really accurate. Deep learning is simply machine learning where you use deep neural networks, which are neural networks with many layers. It doesn't have to be supervised learning, and most reinforcement learning is not deep learning (although there is a subfield called deep reinforcement learning which is part of deep learning).
It’s going to be used more for manipulating users to separate from their cash.
Yep. AI gets easier if you buy micros and impossibly difficult if the game is F2P or you paid "just" $60 for the base game.
Publisher: Finally, we can develop video games without those uppity developers! Okay AI, tell us the magic secrets!
AI: [Make games with compelling gameplay for single players first and foremost and no intention of charging the players extra unless you are developing new content]
Publisher: *throws computer out window*
That industry is mostly dying and getting more and more anti consumer. So, big editors abusing AI won't get us better games, just more and more cloned content that no one even made.
It isn't dying, it's switching to indie. Indie games are awesome and are more alive than ever. And AI would help them tremendously
@@fritt_wastaken I suppose the thinking here is that by leaving a lot of the gamedev process up to automated processes, indie devs would be able to make games with less effort, time and costs?
Looking forward to seeing where this goes, even as the clinate around AI hardens as unethical people figured out how to do it to rip off the skills of artists and claim it as their own, causing a whole anti-AI art backlash, and more and more people coming to see AI in a lot of fields as a job-stealing measure that the wealthy will use to sabotage the workers. In between the many upset voices clamoring for it's downfall and the same tech bro's who once were sooo hype on NFT's now turning their eyes towards AI as their latest grift which they praise endlessly and vapidly, I see this channel as a good way to learn what AI actually DOES in tech... and figure out whether or not I should actually be worried.
I appreciate the support. I feel it's important to stay grounded and be realistic about how this technology works.
Right now there's a lot of noise and hype around AI, and a lot of that noise is specifically about AI for games that is, frankly, unfounded and based on naïve understandings of how game development works (the Web3 bros are coming for the AI now). It's frustrating to watch, and I prefer to focus on what actually works, rather than speculation and hype. As with every video I make, I can point to the source and why it was built like that.
I have a rant of sorts that is brewing in me, that I may well write up and publish on my second channel, given I think we're heading for a new wave of cash-grab nonsense asset flips that are all rely on AI-generated pipelines. Meanwhile there will continue to be smart and effective use of ML in game productions to achieve real results, that still understands the need for having humans in the loop and more traditional design tools.