0:00 hello guys and gals me mudahar and uh while I'm still sick I am absolutely excited to make today's video because ladies and gentlemen remember a while back not a while back maybe a couple weeks ago uh I talked about doom and how it was entirely going to be like AI generated in the sense that there is an actual project where people have generated Doom from thousands of hours of footage of Doom now obviously I think this is pretty wild in terms of you know Computer Engineering AI stuff but the time it wasn't really something that was playable it just generated videos of Doom you might as well just play Doom on an actual computer okay but of course today we're looking at something known as Diamond which is diffusion for World modeling visual Details Matter in Atari
0:46 so we're going to skip back the Atari stuff because ladies and gentlemen this is Counter Strike Global Offensive or counter- strike 2 whatever the whatever the heck it's name nowadays entirely running through artificial intelligence now I wouldn't be making a video on this if all you could do was just generate footage that would be stupid no today we're actually going to be playing our first AI game on YouTube insane stuff so of course ladies and gentlemen one of the clips that I saw over here was this one uh researcher like playing Counter Strike through this artificial intelligence now this is running from my understanding on an RTX 3090 which is still a pretty expensive and very heavy card so today I'm running this on an RTX 4090 underneath my system so the first
1:33 thing that I had to do is I had to clone their GitHub repository and one thing that I will say is generating this GitHub repository and like actually getting it to work came with a whole plethora of Errors there were some consistency problems I literally had to go into their actual requirements folder and just install every single dependency one by one by one which wasn't really the biggest you know pain in the ass the big pain in the ass was getting this to work on my actual graphics card yeah if you're running this without a graphics card you're looking at a frame rate of one one even if you have some of the best processors on the market so a GPU is absolutely required and a very high quality fast GPU is exactly what you need so of course first things first obviously we're like a couple minutes into this video I'm not going to keep you here too long ladies and gentlemen
2:19 we're going to actually fire this up and see AI Counterstrike all right so here I've got the actual thing initialized ladies and gentlemen right over here and we're going to hit enter and uh lo and behold if you're looking at it real carefully ladies and gentlemen uh obviously this is a bit hard to see so I'm going to zoom in through the power of editing and yes ladies and gentlemen this is in fact Counter Strike Global Offensive running in an RTX 490 at a frame rate of eight I would assume if we're lucky and uh yeah it's actually generating the entire world as I'm feeding at actual input so if I hit space jum jump you can see that it recognizes the jump key my character actually jumps in the universe and then
3:06 stay still so yeah this is basically the game World imagining and if you look very carefully it is absolutely um getting a little crazy like if you look at the distance uh it's very easy to break this Matrix and of course the more we jump uh hopefully we're Landing so we're dropping a little bit and uh yeah we actually do touch the floor and uh we are going into the first Arch so right over here and we can go to one of the old bomb s sites you can see like the distance shadowing things are kind of a blurry mess and that's just one of the things with artificial intelligence as it Den noises the image and it looks you know normal this is trying to literally build a world there is no actual game
3:52 code from Counter-Strike Global Offensive this is literally an AI model Imagining the next actual frames as I'm giving it input based on obviously footage that has been gathered about like Counter-Strike dust 2 which is by the way one of the most popular maps on Counter Strike for anybody that has no idea what Counterstrike is so of course the mouse I'm moving around you know if you look very carefully I can like shake the mouse I can shake the shake the mouse move my movement around and like you know change the viewport and obviously one thing that you're seeing over here is it's pretty good like it's actually got the [ __ ] logic down so where it really kind of starts to break a little bit is when you you know move the camera a little bit to the right so you can see the vehicle is over
4:37 there things are starting to blur up and you know if you give it some time it'll kind of stabilize but of course if we turn around a little bit here and there you'll notice okay the actual tunnel's still there so it's kind of got the logic and again this is all built off of the hours of footage it's got from you know all the dust to stuff so yeah you can see that obviously it it it really does work so if we can enter over here you can see that I am now in a different part of the map and of course I can even fire the gun as well so I can hold the I can hold the left Mouse button down and if you kind of look at the gun it does appear to be shooting although it's it's it's it's nothing impressive if I hit three I can switch to the knife or is
5:25 the gun's just going to keep like changing in between like a submachine gun and assault rifle no here it is I got the I got the actual knife and uh yeah so obviously right off the bat I'm going to cut this off right here nothing is super duper impressive but well I mean it is pretty impressive but obviously frame rate Wise Way far from playable so to give you guys a context of this one thing the way that this entire process works is they apparently have trained a diffusion model to predict the next frame of the game so the diffusion model takes into account the agent's action basically as I was you know moving my mouse around hitting WD and then it simulates the actual response so this is one that they have with you know Kung Fu Master so one previous frame and then a current frame so after you hit left and punch it
6:12 should in theory get rid of the actual enemy that you have punched and again you can see that in the process it generates the image and den noises it until you get something that logically makes sense so of course they tried it with a whole bunch of other games like breakout for instance um again I'm not super into like my whole Atari gaming knowledge this is far far far far older than I am but of course you've got you know all these other games you can see like an actual representation of how breakout seems to be working and yet these these are like Atari games now if you look real closely I actually did get these Atari games working underneath my Mac M1 so it is going to be laggy if you look at the footage over here you can see that I'm interacting and moving the paddle at the bottom left and right as is also denoted by the actual programs debug process so
6:59 there is actual like gameplay that is happening based off of again the ai's interpretation of how breakout works now I don't know if this is entirely digging into the game logic or the ROM or if there's some like hybrid emulation going on but this is pretty impressive stuff to see right we're actually dealing with AI generations for video games happening in real time so while this is laggy I want to also again preface this is my M1 Max so it's a bit older when it comes to Mac hardware I'm sure this might be faster on like an M3 processor but if you look at the actual performance you can see that I'm not really getting a lot of you know good frames for the amount of CPU usage which is peaking literally at 100% like on this CPU like
7:45 it is actually throttling the system this is how much this stuff consumes so to go back to The Counter Strike right which is you know what we're looking at here one of the ways to make it faster is to get rid of some of those Den noising steps so can actually simulate it faster so going into the actual configuration files of this uh model which right over here if we go to config trainer. gaml and we just do edit and we go to world model environment and we just change this to fast which should get rid of some of those denoising steps making this a far faster experience this is going to really blow your mind so we're going to hit play right over here and of course it's uh it's firing up it's getting the model ready now you can compile this to make it faster but like
8:30 I said in the beginning there's a lot of dependency errors with this [ __ ] so I couldn't actually get it to compile cuz I needed to do a whole bunch of other things uh that said any software that I show you on my channel as since the beginning um run at your own risk we're going to hit enter and this is where again it blows Minds so if you're looking at this real quickly ladies and gentlemen this is substantially more playable than when we last looked at it so if you're looking at this real carefully this is obviously a version of Counterstrike Global offensives AI that is running at a higher frame rate I think at least we're getting 40 uh something frames and again this is an RTX uh 4090 so again understandably this is the fastest card I think you can get on the market and it still has issues uh
9:17 or it's still not fast enough so of course you can see that my gun keeps changing like camouflages like at one point it changed from like one version of the M4 to another and of course within like a second it just absolutely Came Crashing Down so of course the more we move throughout this environment the more the AI is literally imagining where you could be so here we've completely failed so if we had enter we're in a different environment we've reset everything and of course you can see I've got the oh whoa okay already broke so here it is we're going around um we're entering in the area so yeah we've got the actual um door over there that's open let's get to CT spawn so CT spawn is fine uh so if we go over here
10:03 and uh we make a uh we make a turn on long you can see that long over there is calculated pretty well my gun seems to be completely changing up I'm switching off to Like a Knife which you know here I've got my knife which has turned into a desert eagle now which has turned into a tech 9 now if I go up here you can see that like okay give it a second we are now spawned in a totally different location I'm now long you know people always say like AI generated stuff feels like a dream but this is literally the exact definition of a dream like this feels like I'm actually playing a dream world right now so again no matter where we go my dream takes me back to long it's like [ __ ] Silent Hill in this universe dude I'm like back to the
10:50 beginning so if I try to go up and plant this bomb which I don't really think that oh wait no I might be able to do so I'm right here oh my God I I can actually plant the bomb oh NOP we switched to an entirely different area we are now in a different world model uh okay can't really scope down hit enter okay so this is one that's interesting so if I turn around over here one thing you'll notice is obviously it's broken like the thing about this kind of a model is like it has been built off of so much data from dust 2 like video footage which we'll look at briefly later on just uh down the road as I'm oh I think I'm falling through walls that uh the way that it's figuring out the map for instance like long is that it's seen this Frame like a million times like this vehicle so it
11:36 knows that if I turn right it should be long as it knows that if I'm facing down long from this perspective if I turn 180° it should be either this wall or if I hold right I should be able to go up this ramp so again I don't know why I decided to reload maybe somebody at this point during the recording morning decided to reload and the AI thinks that's just logically what makes sense so that's kind of how this stuff is working at the end of the day like there is no code from valve there is no map data from dust 2 it's literally just taking photos and videos of this map that it's seen and trying to judge location logic uh based on the current perspective that you're seeing now
12:23 obviously this is incredibly impressive because it's happening in literal real time so at some point when this the stuff obviously breaks like here it's just trying to generate garbage frames because it really doesn't know what to go off of like it's trying to make any sense of it and again it's much like dreaming right like when your synapses and your brain fire off I mean you're literally like trying to make sense of random uh you know incoherent things and that's kind of what I feel like is going on with this artificial intelligence so one of the things that people can kind of speculate is obviously down the road with AIS like this what if somebody decided hey what if I fed it like hundreds of thousands of hours of Grand Theft Auto 5 footage right and what if I said okay now generate me a
13:09 new Grand Theft Auto game that I can play entirely through artificial intelligence now obviously I think the complexities are going to be a little bit more intense there but yeah this is one of those like I guess future possible scenarios so to go back to the entire project one thing that they mentioned was like this is the data set they used Counterstrike death match with large scale behavioral cloning so if you open that up over here you can see that this is the footage um right here you can see that this is all the this is like examples of it and I think they mentioned how many hours of footage they needed we used 5,03 episodes to train the model this corresponds to 5 million frames or 87 hours of gameplay you needed 87 hours of dust 2 to generate the world that we did
13:55 and of course for them the training time took them 12 days on an RTX 490 which again is pretty impressive based on the stuff that I've seen so I I want to say like a month ago we went from like looking at Doom AI generated you know not in real time to now playing Counter Strike AI generated in [ __ ] real time the amount of the the rate that this technology is progressing at is actually insane and do I think that this will replace any form of gaming down the road not immediately just in the sense that obviously for a multiplayer game like Counter Strike you need something that can handle World um you need you need an AI that can understand and like keep track of locations in the world
14:41 right like obviously the only way that this was able to really map out locations was you looking at the perspective and inferring uh what should come logically after looking at 87 hours worth of footage you know immediately like turning your camera 180° sometimes just broke the AI because realistically wasn't like calculating a map or anything of that nature so if AI can figure that out interesting and cool stuff and obviously for a future crazy Endeavor like Grand Theft Auto like an open world game it needs to actually remember the entire open world that it generates which I think is a far bigger task than one would even one would even like you know think of but of course you know this stuff is progressing so fast that who knows maybe before GTA 6 we might even [ __ ] get GTA 6 in a die so
15:29 obviously one of the things that I've seen floating around on the internet and I'm sure you guys have seen it too is these kind of videos where people are creating video games but regenerating them with like video tools like Runway AI so there's one video I saw that blew me away like two weeks ago was this one right here a GTA 5 gameplay but it's reimagined by AI now it looks kind of like a a uh Playstation One it almost looks like a old school like PC like click it almost feels like a uh CL pointand click adventure game like watch this it's like an old FMV Style game good to see you man forget our language Ki maybe a little I've been here 10 years you could speak English remember we learned up the English girl
16:15 they be like hold on why did they just change characters uh it's because of the limitations I think of like Runway AI like I think you can only generate like 10c long videos so they have to keep like you know just splicing together the footage but to go forward a little bit beyond this cut scene this part actually blows me away so if you look at it real carefully I'm going to close the music down because obviously copyrighted um soundtrack you can see that it is turning Grand Theft Auto 4's root gameplay footage into apparently a pretty interesting view through Runway AI so obviously it's got the uh rain soaked streets that you find in every Graphics mod but it's got fog it's calculating Al gunin in the background um I mean this is like interesting stuff
17:00 when you got the radar in the corner it just assumes it's like a mirror or something so it's got this in intense blur it almost who's like watching an old school like Russian gangster movie it's insane how good it is and obviously a lot of this has to do with how it's prompted to present this footage but yeah it's it's trying to figure out the objects in the distance the traffic and in some cases I will say it's pretty good with like logic consistency here like trying to keep things looking consistent in a world but yeah to go back to the project over here obviously this is something that has been pushing my RTX 490 as I've been making this video to 100% now obviously very inefficient in terms of like utilizing a graphics card you know you can go from
17:46 you know running cyberpunk 2077 with full path tracing or uh you can run Silent Hill 2 with like the settings completely cranked up and max out that GPU or you can run an AI dream version of Counterstrike Global Offensive now the thing is obviously you know down the road you know a year from now I want to see where this kind of Technology really goes from like are we actually going to have that Doom completely playable within like this AI model that would be interesting to see but yeah when I saw this I was kind of Blown Away I thought maybe it was just generating videos to begin with but to actually be able to play my first AI game wild [ __ ] indeed so if you like what you saw please like comment and subscribe dislike if you dislike it I him out
