Why Doom? Because it's been ported to just about every digital platform in existence, including fridges, microwaves, watches and more. If it has a screen, someone has ported doom to it.
It runs on my rc controller for my fpv drones too. Recently bought an Quest 2 and it is awesome playing the original Doom and Hexen and Quake in fully surrounded VR.
@@staomruel Yeah, though technically, I believe the pregnancy test was "upgraded" (as in they changed the CPU to an equivalent CPU because the old one couldn't be reprogrammed and changed the LCD screen because the old one could only show specific things). Or at least that's what I heard... But it's still impressive af, especially since iirc the new CPU wasn't more powerful, just reprogrammable.
It’s a meta joke. Doom was about technology leading us to confront the paranormal/spiritual/other dimensional, and now the irony of AI brute forcing Doom into existence is perfect.
@@ThompsonDB brother, observing the current worlds events, I am convinced that biblical end times are upon us... but it's not exactly what bible believers may expect... a prophecy is man-made, and man-fulfilled, by those of the same bloodline that betrayed and killed God... I do not understand their reasons, but it's not for good, and bible is not what it's supposed to be.
It's like a dream. Doomguy looked at the hall, there were armour pickups. He looked around, and the armour pickups weren't there anymore when he looked back at the hall. It happens if something catches my attention in a dream.
Everything I've seen from ai generators feels like a dream. All of the weird "hallucinations" in pictures and conversations remind me of my disjointed illogical dreams. The two best uses for ai-gen is uncanny-valley nightmare-fuel, and memes.
I'm still more interested in AI run NPCs and dense populations. It's jarring seeing "ghost towns" with NPCs in a loop, games are becoming more "realistic" but that stuff is still immersion breaking.
This kind of thing is an important part of that. Being able to generate new content randomized on the fly will massively increase the lived feeling of games.
Hell ya, and NPC's leveling themselves up and creating their own agendas, and time in the game world still moves on and NPC's evolve even when your not playing the game..
I remember when Wolfenstein came out, then Doom, and Doom II. Man, they were so unbelievably ground breaking at the time. Doom II was the absolute GOAT.. so epic.
Track down Masters of Doom (audio book version is excellent) as it's the story of Carmack and Romero from commander keen through to Quake and how it all started. It's more wild and crazy than you would expect - there are thefts, fast cars, other big studios from the day make cameos it's so good. Kinda surprised it hasn't been turned into a movie or tv mini series yet.
For me it was the Quake "test" release. True 3D environment. You could look straight up, straight down with no distortions. Blew my mind. Ran great on a Pentium 90.
Haha- True enough. That said, I think the new TPUs are a LOT more energy efficient. The previous models were running at 200w/hr, while the new Edge ones are at 2w/hr. I don't know if this was run on the newer TPUs, but on the plus side, it looks like we'll be seeing a pretty big reduction in power output as the newer generation rolls out. Believe me, it is something that I think about as well.
I bet they don't work anymore. Granted I haven't used floppy drives of that kind since probably the 90s or POSSIBLY early 2ks, but I seem to recall that if you left em sitting for long enough they wouldn't be readable when you went to use em.
I imagine, like procedural generation, this kind of tech will be handy at "filling out a world", but game developers will only end up using it in a piecemeal way cuz investors love their buzzwords, but in the mean time, you'll have curated "story zones", and then fairly random meaningless Ai-gen zones.
Oh, I for sure see that. But I can also see the modding community picking up on this. That's one of those things I always wished got bigger for a game like Skyrim-- kind of like old paper and pen RPG modules. That might be kinda cool.
DOOM was the greatest game of all time... until DOOM 2 came out and then DOOM II was the greatest game of all time and I haven't seen anything since that's dropped as many jaws
I still can't watch the movies 'Aliens' or 'Evil Dead II' without immediately remembering all the mods in DOOM - getting my hands on that MOD was one of the greatest times... "Let's Rock!"
it's really freaky to me how the AI seems to add tiny little details that are just SLIGHTLY off, but not enough to be noticeable unless you're scrutinizing. things like, the acid being a little wiggly, or bullet holes phasing on and off of walls, or the two points where walls meet being a little messy. also, the way items and barrels seem to appear and disappear when not in view. cool stuff, but man there's something so unsettling to me about seeing AI make mistakes like this. the early footage where the monsters would melt behind the wall made my skin crawl. I think it freaks me out because it feels like, at any second, the entire simulation could fall apart and suddenly you're being shown these horrific AI-generated amalgamations.
The paper keeps using the term "simulation" so it's more generating a video of the experience of playing Doom. A little like the shadows in Plato's cave.
@@SimplCup well if it currently generate at 20 fps on a modern computer images of a game that was made to run in 320x200 you can guess it is not efficient and will not be.
@@magen6233 well obviously it's not yet, cus they just experimented lol. They used really inefficient sd 1.4 and somehow made that architecture generate 20 fps, while there are already architectures that are way faster and higher quality than sd 1.4 like sdxl turbo for example. With sdxl turbo you can generate images of higher quality really fast probably around 5 per second if you just use the model, but I'm sure if they made slow and old 1.4 generate 20 fps they can easily make sdxl turbo generate like 300 fps or even more. The progress of ai in this 3 years is just insane, so i can't even imagine what ml models and architectures people will be able to make in 20-30 years.
If you think about it, you're essentially playing a dream or nightmare(since it's doom), where the spatial surroundings change with no continuity. Btw. I love the breakdown of the technical stuff. Keep it up.
In the DOOM storyline, the UAC research facilities slowly get transformed by hellish influences as you progress. This could be very effective showing that effect in real-time.
@4thlord51 Actually, it's even better than that, it has like 90% continuity, making the changes unsettling rather than just destroying any and all sense of immersion.
I already hate this project but I want to be really clear, the RL agent and I assume also some of the visual portions of this are trained on a purpose-built fork of ZDoom, which is *not* technically original Doom. that's why there's randomly appearing/disappearing bulletholes and smudgy sprites instead of crisp pixels, there's a texture filter that is defaulted to "on" and I assume Vizidoom did not turn this off for its training model. also found it really interesting you didn't mention the human identification portion of the paper :)
@@EldenLord-wh4ov Not even an emulator. It is more like a drive through google maps with occasional opportunity to feed a cat or something (which would become unfed again once you turn away, because no state is being tracked).
I never knew this. Spent 20 years getting irritated with people getting it wrong then kept hearing people say it right. Then people like you confirming!
@@l30n.marin3r0 john carmack and everyone at id called it id like as in kid in all the video footage of them. no one from that company has ever pronounced it different. As others have said it stands for ideas from the deep. They decided it was too long so called it id. and they themselves said they likened it to Freuds' id because that is the pleasure center. So the ai probably should learn from humans like him
I think this is obviously a development dead-end, but it's neat that they got something like this working. I remember a long time thinking about a theoretical game that was produced with trillions or more pre-rendered or painted images. Like if you could paint every possible combination of pixels in the game and then the engine would just serve the correct image then GPU hardware wouldn't be a limitation and you could have graphics that are as advanced as real life, and basically be photo quality. This is kind of that idea but using ai to generate the images instead of having to make them all in advance and store them which is obviously impossible, besides that the logic of knowing what image to serve would also be impossible. I doubt this type of technology will ever be used in the lifetime of anyone alive today unless there are totally new technologies that develop over the next 50 years that make this practical, but there is no reason at the moment to think that will happen. The limitations of todays LLM technology mean that no matter how advanced they get, no matter how fast they are, this will never work because they are just way too unreliable. Even if they somehow got 100x more reliable which is probably not going to happen, they would still be far too unreliable. Same problem that makes quantum computing probably useless for most tasks. You need literally 100% reliability from hardware, or the core of the software, for most things to work. 99.99% isn't good enough and AI is at like 30-50% accuracy for lots of things. Going from 99% accuracy to 99.9% accuracy is a 100x improvement, same as going from 99.9% to 99.99%, so as it gets more accurate it will become exponentially more difficult to make noticeable improvements. Modern games are many orders of magnitude more complicated than DOOM, they had to train it on images from the already completed game, the hardware is something that people will probably never own and is tremendously inefficient. It's a neat POC, but definitely will never go anywhere. Honestly don't really even know why they would bother building this out besides to attract investment from tech illiterate investors.
The tech is impressive from a technical point of view. But this can't create real games until the context is multiple orders of magnitude bigger. This "game" can't remember what enemies you killed. It can't remember how the room looks behind you. You don't have a map; anything that is not on the screen is destroyed and replaced with something new next time you check. This limitation doesn't look like it's going away, and it's a total deal breaker. This can only be used for games where basically all relevant information is on the screen at hall times. So like, Tetris.
Would make for a great horror game though. Reminds me of a recent title called 'Layers of Fear'. The gimmick in that game is that as you look around a room, the game itself will actively change the room you're standing in. As in, you look north, you see a wall. You look East, West, South, etc: just more walls. You're hemmed in. There's no escape. You're stuck here for all eternity.... or are you? You look north again, and now there's a door in the wall. A creepy, ominous, decaying door with blood on the handle and strange sigils on the hinges. Do you open it? You turn around and look south again, examining your options before you make a descision. Now, there's an enormous mass of writhing flesh and flailing limbs; wailing, pleading, lurching toward you. You know the game will likley de-spawn the door in the next 3 seconds. Do you go through that door?
Won’t even work for Tetris because you need a storage array to store the score… Basically anytime you need to store a memory of anything it’ll get corrupted so this thing isn’t making game dev obsolete for a long time… Not to mention, it’ll just put all western devs outta work but japan won’t use it… Good thing iam moving there…
This is more of a proof-of-concept to demonstrate that they can conjure the game state in real-time. To make a complete game, it's either going to have to keep a running model of the entire game world, or have it render out finished .WAD files.
So, less like playing Doom, and more like experiencing the dreams you have after you’ve played too much Doom before going to bed (yes, I’ve done that. Still have the occasional video game dream, though I think mine look more like Unreal than Doom).
@@jasonlescalleet5611 good old unreal tournament 😊. Brings back good times If any AI can simulate that kind of hectic gameplay I'll be very impressed. There's always so much going on on screen at anyone time - the AI would have to actually understand why bunny hoping works, the ballistic trajectory of a rocket fired off screen, the fact that health packs exist and respawn on a timer, etc, etc We'll know humanity's truly doomed when it can simulate all that without frying the servers 😋
Wow! It can, while utilizing some of the most expensive and cutting edge technology in the world, barely hit 20FPS while poorly replicating a (by modern standards of technology) extremely simplistic and primitive game that is over 30 years old! This cHaNgEs eVeRytHiNg!
It's pretty cool that they managed this, but as with all of the Generative models atm, you cant really use this to make anything new. And even more limited with this, in that all it can really do is copy a game 1:1 at best, but will most likely just introduce bugs and artifacting.
Very true, but it’s a fairly big step. Right now; we’re looking at kinda janky mods for games. But down the line I can see having a RL Agent learn a game style, and then be able to generate a different game in another style. Granted, you won’t be making anything new- but neither has Ubisoft! Zing!
Generative models CAN generate something NEW.. Create something NEW from what already exists.. the only source of new stuff in this universe is FROM CHANGING STUFF THAT ALREADY EXISTS..
Hold on a second... When Doom released, it was running at 24FPS. That was vanilla Doom's sweet spot. So this thing is doing it at 4 frames a second slower than when it was brand new. Considering the radical departure from your standard 486DX2@66MHz PC Doom machine from 199X that's very, very impressive.
486 66 was pretty high end at the time. Most people still had 386s. It’s not like the TH-cam era of today where everyone had the latest and greatest of everything
🎯 Key points for quick navigation: 00:00:00 *📜 The history and impact of original Doom* - Introduction to the original Doom released in 1993, - Explanation of how Doom was groundbreaking with its 2.5D engine, - Discussion on Doom's continued influence through WAD files and modding. 00:02:59 *🔍 Introduction to the AI-generated game engine* - Explanation of the first AI-powered game engine developed by Google DeepMind, - Differentiation between usual AI-generated videos and playable AI-generation. 00:04:35 *🤖 Technical composition of the AI game engine* - The game engine combines an RL agent and a diffusion model for AI gameplay, - Reinforced learning allows AI to understand and play Doom. 00:06:10 *🌍 Understanding World Models in AI* - World Models enable precise understanding and anticipation of game events, - Distinction between AI-generated environments and procedural generation. 00:07:44 *✨ Challenges with AI gameplay consistency* - Limitations of the AI to maintain visual and gameplay consistency, - Degradation issues during human interaction not noted with the AI agent. 00:09:11 *🎮 AI Doom's potential and future in gaming* - AI's impact on traditional game development and possible new branches, - Current state limitations and future possibilities with scalable AI game engines. 00:10:21 *🎥 Why Doom was chosen for this AI research* - Doom's technological legacy in 3D graphics makes it ideal for AI research, - Its rich historical impact and complexity offer a challenging AI test bed. Made with HARPA AI
I've grown fairly skeptical about the future of this type of AI. The usual line is "imagine how it will look a few years from now," but big limitations are becoming apparent both in the methodology and in the availability and quality of training data. It seems clear to me an upper limit exists for this technology and we may be already bumping against it. I don't think it'll pass us by as a fad, but it also won't take over every field like all the hype keeps insisting.
A game that came out around the same time was Mechwarrior 2. It's a 3D simulator but had awesome graphics and was technically impressive for the time. I always kinda lump that in with games like Doom from the time
Oh, I LOVED that game! Apparently there was a bunch of rights issues that killed the franchise- which is a shame, considering how awesome that world was.
Cool tech demo but seems to he reinventing a worse wheel. It makes more sense to me to use ai as post process effect. Traditionally render a basic structure of topologies and scene elements. And have ai convert each frame into style of your choice. Which there has also been some research demos done using that idea.
It opens the door way more than that. As to say, you more than likely will still use current engines and game design, but allow the AI to generate all sorts of effects either post or on demand. For example, a real frame of the game when you move the mouse/keyboard and the real game engine is taking way too long. Instead of tring to creating a frame base off visual elements, it could be base off the game/engine it self allowing for lower input latency. Another is the fact you could add in features that take less over all elemental power from hardware such as RT that could be part of the post process.
Maybe I'm archaic in my thinking about gmaing but I find this to be the apotheosis of video game design. I love being able to talk to another fan of a game and understand that we share an experience. This seems like the next level of procedural generation, where players won't really be playing the same game at all. Everyone just playing the own divergent microcosm of many games. Less and less intent and game design from the creators. On the bright side, it provides nearly infinite possibilities for content creators.
Nope. I noticed on at least two occasions in clips shown early in this video that collectable items vanish when you turn around without picking them up! That's a serious glitch in a game like Doom, especially when some of those items are ammo. And I wish people would stop calling these machine learning algorithms AI, they are NOT AI. They are just more advanced versions of algorithms that have been around since the 80's at least. I have seen them on the Sinclair Spectrum computer back then and they keep popping up through computing history with the promise of being AI but never delivering. The only difference from back then to now is the power of computers and the amounts of storage and RAM you can throw at them to make these algorithms far more effective. One SciFi story I read recently has the right idea. They had actual AI entities that worked with humans and an enemy that was thought to be AI but turned out to be a lesser version called VI, Virtual Intelligence. The difference being AI can actually think for itself with original thoughts and be creative but VI can only mimic intelligence. It can respond as if a real intelligence, copy what it has seen, and pull from knowledge it has been taught. But it cannot create, it cannot think for itself, it is not autonomous. This is the same as our AI today, it's not AI it's VI.
Oh, 100% with you on ML vs AI- but, I think Apple pretty much proved that no one (for some reason) is willing to call it ML. They doubled down hard on the ML term at the start of all this, and while (to be fair) there were a lot of reasons they lagged in the AI race, I don’t think their insistence to use the correct term helped them at all. What’s the name of that story? Sound interesting!
@@TheoreticallyMedia It's an audio book on TH-cam, called Terran Contact on the channel SciFi Stories. Also, I think the term VI is better than ML. I think people would be more willing to refer to it as Virtual Intelligence.
Whoops, wrong story title in my other reply. The actual story is The New Species. Same TH-cam channel as I said before. I could have edited the comment but you would get notified of the change lol
So correct me if I'm wrong, the game is generated while you play it. Then you can prompt for example a "Classy Doom" where they throw champagne bottles, salads and lobsters at you wearing tuxedos and silk dresses or a "Laundry Doom" where they use washing powder, softeners and even whole washing machines to kill you? Correct?
Omni, play The Sopranos in style of Simpsons, sort Google Doc folder by Date Modified,. Build me a shitty resume season with Bart and maybe that Wendesday girl I keep hearing about on Civic. Can Paulie have a tail? Run all audio through the DonaldDuckHD.onnx model and pitch it up by 12. Oh, and no clothes. on anybody. (pulls shades down and looks into camera) GOOGLE cancel everything I just said and just build me a world full of nude Lorraine Braccos and Edie Falcos. Helllo? (but something goes wrong and it's just Lois Griffifth from Family Guy)
Ripping through fossil fuels with a ludicrous amount of compute to generate each frame just to end up with a mediocre game that wasnt designed by a human artist with anything to say... Everything about this concept is terrible
If by "change" you mean "chalk up one more thing that used to be good until a flood of execrable AI garbage tainted even what was left of my fond memories" then yes, it does Change Games Forever.
but what is the output? As fact, doom map generators is exist since ~2014-2015, but they they were not popular at all. Because maps is fun to make. There is another biggest problem - character animations. AI can't make high-quality sprite animation of low-resolution characters like doom or classic fallout, and this is the main problem in indie game development, where mocap is not available. And obviously no one will deal with this problem, AI will always be designed for ultra HD... So, nothing changes - map generators for uploading textures have existed for a long time already
It's impressive tech... I just don't see a point to it. I much rather they worked on an AI that can generate proper enemy behaviours instead of wobbly bullshit graphical elements...
For sure-- but I think research like this are all the little steps involved in eventually getting better at the really good AI stuff that will be useful to game devs.
I dont think this game world generation will go anywhere, but i excpect developers will stop focusing on graphics and only one gameplay, level design and animation, essentially. The AI would then redraw the game visuals based on crude underlying graphics. Essentially the game engine would produce a controlnet and the ai could produce varying visual styles based on that. But AI image generation would need to become a lot faster and cheaper before that.
I understand now. So it's soul destroying. "Hey, what skill set can we replace with AI this month?.. I know.. game making... if...if we could let anyone create a game then everyone will want everyone else to play THEIR game and then nobody will play games anymore because they're all bored of all the games... wow.. we can do that and have them all make AI textures, AI 3D models with AI voices and AI sound FX.... so... we've destroyed the point of human creativity and uniqueness and eventually AI will have nothing new to give and people will have become so expectant of instant results that the idea of creating anything on their own will become a futile chore because someone or some machine has already done it..."
@@Mewseeker Agreed. It's the movie "Idiocracy" in the making where all the intelligent and creative people with modest intention and morals die out as the dumb spew out kids and AI garbage and become heavily reliant on a single mega corporation that bought up everything and decides what people eat, drink and love. Ultimately a few selfish biillionaires with no concept of life (like NVIDIAs boss) caring for nothing other than bragging their own genius whilst claiming its for the world's benefit! Grrr...
Now we only need ppl start making "great games" which get popular and not DEI virtue signaling crap which flops in the first few months. Same situation as with art.
"if we could let anyone create a game then everyone will want everyone else to play their game" You're grossly underestimating human laziness. For instance, everyone online has access to the culmination of human knowledge, literally at their fingertips, but people are now dumber than at any point in modern history. _Ironically, psychologists have noted a steady decline in our collective intelligence since the birth of the Internet._ In short, the vast majority of people will lack the initiative to make full use of the tech. When AI gets to that level, most people will just make memes, and sh*tpost with it. Only a small percentage will use it to its full potential.
Nearly shit my pants when I first saw the paper. This is actually kind of a massive breakthrough, I’m genuinely curious to see how well this tech scales
@@MittzysI believe they meant it as a breakthrough in AI development, not game development, although it could lead to breakthroughs in the gaming industry. It would be like an overpowered game master, making the game as you play it in real time, adapting to your decisions. I think it would be especially fun if played with other people, like D&D. A human DM could also be present to further influence the AI. This could open a lot of possibilities for individual players. I think it would just be another way to have fun. People would still make non-AI games.
@@shizumi5243 « It would be like an overpowered gamemaster… » Yeah, that’s basically what a human game master is and does. Why wanting to put him/her out of the equation and get replaced by an AI ?
@@Mittzys brother I work in game development. I have a B.S. in Computer Science, and in my professional opinion, this is actually a really massive breakthrough. There are tons of cool ways that *actual programmers* could actually use this technology (granted, in a few years at least) that would make video games better. Lots of developers that I know are pretty hype about this project, and it’s kinda disappointing that instead of trying to have a conversation about it, you immediately decided to insult someone’s credibility with no basis.
Not 100%, as to say, it opens the door for different way we can make hardware to run whatever features we want to display in a game. From everything like RT to the game it self using a more dense design than today's hardware or for at least on demand frames.
@@gabrielandy9272 More or less... is not an AI... which means that right now, is "imagining" stuff due to the training it had with the actual doom. This means that at some point, even with training of all games that exists, when you extend it powers to generate random levels with the same pattern, the series would basically converge. You can argue that "well, but all levels would be 'different' even if they are the same from a high level perspective", we could assume that that might happen (they will clearly repeat but it's mathematically irrelevant how often that will happen or even if we can see it in a lifetime of playing). But then, that means that we already have that and it's called Daggerfall. Infinite dungeon crawler and questing experience! All of them are different! And if you are thinking of that type of infinite we can totally code that already and probably make it run on a toaster for doom. Given that the toaster has a display... or... maybe it could print the frame in the toast... So, as for right now, this is just a step up of the technology that already exists that generates videos out of a prompt. But has the potential of being interactive and assuming the demo is not "slightly edited" like we've seen in the past with some models, much faster since it goes in real time and when, could go forever.
This is very fascinating. I'll probably just go ahead and mark my own words that I don't think this is really going to go anywhere as far as companies selling GPUs that will generate instead of render. If it ends up that way and we get insane visuals with low energy cost and consistent framer rates then I'm all for it. As it should be noted, though, all of these things are well outside of my knowledge ability.
@@christopherpekel6096nah, its a reboot. Doom 2016 is the updated experience of playing doom 2 1000 times and wanting more challenges. Doom 3 is the actual remake of doom trying to put more realism and effects into the survival/horror action game. There's more fans who want action over horror it seems...
@@gesi7072 a reboot is when the original concept is reimagined, with core elements retained. Slayer is a new character, the lore is reimagined etc. To be an actual sequel would require the same gameplay, characters etc and overall setting of the originals. Doom 3 is also a reboot of the original. I don't think people understand what reboot means. They can still link slayer to the older games, but its a different continuity.
nope, why do we need low fps, hallucinating DOOM while we already have Brutal DOOM with HD textures and dynamic lighting running 60+ fps on high resolution? Admit it already, large models does not work for what you are trying to use it.
I mean that AI didn't make doom from scratch, it needed the existing massive data set of doom to train on. It super cool but in its current form it doesn't mean much imo until it can make games from a collection of multiple games and mixing the styles and genres.
Counter Point: Companies/Publishers like Ubisoft will /are absolutely be training AI to generate entire "open" worlds based on their own previous IP, and eventually even the points of interest/quests/missions. Specifically, if for nothing else but to significantly reduce the cost of level development.
@@G4M5T3R No publishers they are using ai make the entire game, they use it to help generate specific assets, not the entire game, big difference. They are not wasting money on something like Nvida did here, a game engine for a game that is 30 years old gets worse the more you play it.... Nvidia has cash to burn on research these development studios/publishers, don't. This toy project had to cost a few mill with zero return.. Fundamentally hallucinations are baked into generative AI with the current approach. Untli they can solve the hallucinations these problems will persist. Unreal's gen AI solutions are far more practical and valuable than this research demo. Generating an entire game on the fly then testing it, looking for flaws, nightmare. Generating an asset and then looking for flaws great solution.
@@G4M5T3R Generating an open world is not generating a game. They can already generate worlds, without the massive cost of models. No mans sky for example. I think there is confusion around the word generate, the game is being generated on the fly by the Nvidia model vs generating a world during development. Very different things. Nvidia example here is generating the next frame in real-time. Its a novel experiment.
@@G4M5T3R I guarantee you that any dev shop could build doom cheaper than the process used here and with better results. You are assuming it was somehow cheap.
I still have a hard time thinking about ANY of this as actually AI. Something had to brute force learn how to play doom to figure out what was allowed and what was not. Something had to copy the graphics & characters from the original game. It had to be cajoled into behaving with the stats at bottom of screen. And then it what...it just reproduces the world & elements that it's copied from the original, based on inputs and actions from the player? Is this not just some fancy way of using machine learning to copy an existing game the long way round?
To be fair, it's not even copying the game so it fails even at that. It just imitates it. What you get is frames, but no enemies, textures, doors, places, variables, states, points, etc are stored anywhere.
@@NostalgicPlayYT imagine the crash remake lol : 1)can't hallucinate 2)no blurry wobbly images 3)runs locally on a toaster (compared to the supercomputer needed here) 4)the only training needed is for the player. 5) relatively low effort , hig results . 6) fine tuning and control on the overall result. 7) don't require a roket scientist to play it.
I don't think AI can ever make a proper game. There are too many smaller details from a game design perspective (jump height to platform design, secrets, how do you get out of a stuck area, etc) that AI can't properly solve in a cohesive manner. I think the concept is interesting, but a human mind is still needed to thoroughly fix all of those flaws
For sure. I think that an interesting future will be when anyone can create the game they want to play- but I guess the thing that isn’t in the sales material is that everyone will also suddenly become QC!
I think that if you were unable to accurately envision the 'impossibilities' which have become not only possible, but almost-mundane in the past year alone-i.e., your name doesn't end with "Karpathy" or "Sutskever"-then to impose that same normalcy-bias on future potential constitutes a failure of introspection. If you wish to remedy this metacognitive blind-spot that seemingly 98% of humanity, and certainly most TH-cam commenters apparently share (which, assuming you value your sanity, I strongly recommend doing), I would suggest Eliezer Yudkowsky's seminal essay "There Is No Fire Alarm for AGI".
I liked Nvidia's approach to this. They basically use a traditional game, gta 5 in this study, as the basis and then use AI to generate the final graphics. In theory this would allow developers in the future to make a game with very basic models and no textures and than just assign prompts to these simple models like "car" or "tree" You can then play this game carefully crafted by the devs while the graphics just get generated on the fly.
The only thing an AI will never be able to do is really feel emotions or reach sentience. everything else an AI will eventually be able to do, because what humans do isn't that impressive.
Does that mean in future, Some games will have options to run on Tenser Cores or CUDA cores completely? As far as i know, CUDA is faster than the GPU itself, so how far this can go in the future will be astonishing!
I don't think there's even an ounce of practicality to this POC. This is just a "is this possible?" project, and nothing more. The hardware one is going to need to have at home to run the most recent games of the era is ridiculous. So unless there's a *massive* technological breakthrough that suddenly makes quantum computers affordable to the average Joe, I don't see this concept beating game engines, ever.
A mod for Skyrim adds an AI companion. That NPC's dialog is generated in real-time, depending upon how the player responds, asks non-scripted questions, etc. Another AI mod allows non-companion NPCs to respond to the player in real-time as well. Yes, real-time, AI gaming is coming
I really hope not. If using ai slop in games will become a "norm (the soulless corporate gobshites' success at destroying another subgenre of HUMAN art. Since only real HUMAN beings with thoughts and feeling can and always WILL create ART)" in the future, I just won't be playing future games, lmao.
Let’s use a thousand servers consuming a gigawatt to produce the same result that a single 386SX could turn out in 1993. AI is so amazing! Stock price up and to the right!
The statement "games in the future will be generated, not rendered," is nothing more than a tech-bro catchphrase for people who have no idea what making video games is like.
No, once ai and hardware is advanced enough games will be randomly generated or generated based on the users prompt on the fly. With that being said that doesn’t mean conventional game development won’t exist, but ai will still play a massive role in that as well, the endgame is eventually ai just making its own games in a matter of seconds that would take an entire team of people years to do.
@@DoritosBurger More that there will be an established engine, assets, etc. for the game that an AI will then use to randomly generate things. Because a game trying to simulate everything from the ground up is just asking to crash randomly at the tiniest of problems cropping up. AI isn't and can't be perfect because it's not a thinking entity that can self evaluate to find errors. That's why even when you're asking something like Chat GPT while it can give you a textbook answer, because it's copying a bunch of textbooks, if at any point there's a deviation anywhere it can quickly break down into gibberish. Just look at all the silly videos on youtube where AI bots on stuff like Google give wrong our outright destructive answers, now apply that to a game where something destructive could cause a game to crash, saves to corrupt, or even blue screening a computer. Until AI can self evaluate to find errors and adjust itself that's always going to be an issuee. The issue with self adjusting AI is that's a hop and a skip away from self determining AI, which is much, much bigger problem.
@@DoritosBurger And like I said, an AI will *always* have this problem unless they are smart enough to fix the problem. If an AI is smart enough to fix it's own code it's probably smart enough to go beyond the bounds of it's own code, and then we get into the proverbial paperclip AI territory unless protections are in place. Further, an AI generating an entire games engine and assets on the fly is incredibly resource intensive and wasteful. Even 50 to 100 years from now that kind of generation is going to be resource intensive by default. Having the generative AI built within the engine rather than the engine being generated by the AI makes the systems run by the AI an order of magnitude less resource intensive. It would also let them have a memory longer than the 3 seconds in this video without ballooning out into ridiculous storage requirements. Just look at procedurally generated games like minecraft, it's not all that resource intensive and the world isn't all that hyper detailed, but I've seen world saves that can get up to 2-3 GB in size. Now expand that out into a more high fidelity game with persistent assets like wholly unique towns and NPC's where the game is keeping track of that information and you could feasibly have a game save that gets into the hundreds of GB to even several TB's of space, for a single save. Unless that 4 dimensional crystal storage tech I heard about the other day becomes viable for home use in the next couple decades our storage capability is starting to slow down, it's not making jumps in leaps and bounds but in little steps here and there, and those kinds of file sizes, never mind the power and cooling requirements to run those kinds of systems starts to become infeasible.
As a gamer and game programmer, I believe this is not necessary because experts need to put more efforts to other aspects. Procedural generation is great, and still can be made greater. In same time, gamers need better AI as opponent, or for cooperation. What I mean better is not something like god-like AI, but human-like AI. Funny most games now have great graphics, physics, etc but cant have great AI.
People need to realize that AI isn't smart but incredibly dumb, as is also expressed by the huge computing effort needed to achieve results that then human intelligence has to repair. We wouldn't feel the need for AI if humankind wasn't dumbed down in its abilities to shape smart algorithms, because what we mostly do is offload neural networking in a brute force style.
So a really crappy and slow version of classic DOOM, where there's no intentional map design and your ammo counters are flickering between random numbers. Sounds more like a DOOM-inspired nightmare to me than an actual game.
@@pw6002 Lack of talent? What you on about? Don't you know it's a law that Doom has to be ported to everything? It's true, it's mankind's duty to put Doom on everything with a screen. Just check the UN charter, it's there, trust me bro.
@@the-answer-is-42 I was talking about the lack of talent that is inherent to AI tech bros, not about the fact that DOOM indeed is ported on every existing electronic device in existence ! ;)
@@pw6002 And I was making a joke because I was in a silly mood. I agree some of this AI stuff lack talent, though, but I'll let AI Doom slide for the memes.
what a way to overcomplicate things unnecessarily. it takes exponentially more calculations and provided information to render something *_to specification_* starting from a blank slate, than it takes from a pre-defined model. which is exactly how Doom was designed: "Render [ceiling] from [layout] and [information about height]" if you want the engine to render something it has no idea how to render, you have to TEACH it (AKA "program" it!) how to render it! which you'd be doing WHEN you actually program it! if you MUST correct the engine to render what you WANT to render, accurately and to specification, might as well write it once, set in stone, and direct the engine to IT every time. WHICH IS WHAT IS DONE ACROSS THE BOARD, BECAUSE IT *IS* THE MOST EFFICIENT WAY TO DO IT!
Exactly. A lazy programmer is a good programmer because they learn to optimize everything to save themselves the extra effort and headaches they'd have in the future. This application of AI is a deadend.
Not seeing it happening. And I hope it doesn't.I mean, why using a defective by default technology for an application that is much better solved with traditional approaches?
you're missing the point. first of all, this is only the first generation of the engine. 2nd, imagine your MOST favorite game ever that will NEVER end. 3rd, think how easy it will be to create (or generate) a whole new AAA game designed exactly like you told it, and now you can sell it and make money. EVERYBODY will be able to create full Triple A games with the highest qualities and the best stories.
WOW you made a game that takes next to no resources to play and made it take 10,000,000% more processing power, good job 😆 For real though, of course this is technically super cool and impressive and I'm very excited for the future of GameNGen!
It's amazing seeing this video after something that was shared with me a month ago. I've been coding for 26 years now, mostly focused in AI and networking, so I have lots of friends that specialize in those areas. Last month a few friends of mine showed me this project they've been working on. They started by showing me this FPS with characters made up of about the same geometric density of a Minecraft character and the backgrounds was pretty much flat, untextured, polygonal primitives. So I'm like, "What is this? You guys making some kind of retro Sega Saturn FPS?" My friend fires up the generator engine and suddenly the rendering looks like some kind of hyper-realistic film from a first person perspective. It was running around 5-10 fps, but it was fully functional. I was just standing there looking lost as I couldn't believe they managed to optimize all that data to the point that the game was reasonably playable. My friend says to me, "Remember when you kind of postulated the idea that one day we'd use AI to generate content for game backgrounds instead of rendering vast distances? Yea, we're actually able to do it for everything just really slow right now." I was like, "Yea, I didn't see anything like this coming for a while though." He smirked at me and said, "Imagine where it'll be in 5-10 years." I can only imagine what gaming experiences will be like for my kids in the coming decades.
Nope, this is bs. Plus it was done by Google so most def this was faked for a presentation. Longer version - Im a game developer (20+ years) and also work with AI. AI in it's current form is a glorified autocompleter. In order for AI to produce a game like in the video, you would have to feed it (machine learning) with the original source code of Doom, which is most likely what Google did since it's open source now). If you feed it with the source code, producing Doom levels is trivial - it also loses the impact of what this video is trying to convey "making doom with prompts". RIght, feed AI with the source code of any game and spend countless hours "teaching the AI to autocomplete stuff" and then for sure you can generate levels of any game - that you fed the source code into. TDLR the video is bs, coming from Google, not the youtuber.
I was just reading the Google paper on this - still sounds like BS to me from the same company spilling bs about their non existing quantum computer. But oh well :-)
@@yohamiso what you are saying is that your original reaction was wrong after reading the actual paper. But you stick to your original opinion purely based on that it is Google? Might want to relax a bit there and maybe accept that sometimes innovation and real new things can come frome big companies. They are still filled with actual people you know.
What long winded unwatchable crap. Im still searching for video of gameplay. This is not the sistine chapel its a video game, stop making this so precious.
I like that the last word of the video is actually your name. Most of the people start with their name first and who they are first. I like your "approach" of introducing yourself.
I would say this is the dark ages of Doom mods where people stop creating and instead made an AI make it poorly for them. Tho I know 90% of mod makers will stick with doing it themselves.
"Your scientists were so preoccupied with whether or not it could run Doom, that they never stopped to think if it 'should' run Doom."
I HEARD SOMETHING LIKE THIS BEFORE BUT I FORGOT WHAT IT WAS
"Rendering, rendering is on the verge of extinction. If I were to create a flock of generating on this island you wouldn't have anything to say"
@@philosoFreedomGaming I remember now thank you
"Yes. But. Can your AI hallucinate Doom?"
Doom must run on everything infact i got doom to run on foot fungus
2050: So we now present this holoprojection room with doom running on it...
You ARE DoomGuy! I'm there for it!
VR AI coming to a 5090 near you.
You can run Doom in VR right now, it's amazing.
25000: so we present Doom running on Proxima Centauri's starspots
@@TheoreticallyMedia we present vr doom 1
Why Doom? Because it's been ported to just about every digital platform in existence, including fridges, microwaves, watches and more. If it has a screen, someone has ported doom to it.
Exactly what I was thinking. The real question is why *not* Doom? It’s the most relevant game to test IMO
Calculators
It runs on my rc controller for my fpv drones too. Recently bought an Quest 2 and it is awesome playing the original Doom and Hexen and Quake in fully surrounded VR.
I've seen it on a pregnancy test
@@staomruel Yeah, though technically, I believe the pregnancy test was "upgraded" (as in they changed the CPU to an equivalent CPU because the old one couldn't be reprogrammed and changed the LCD screen because the old one could only show specific things). Or at least that's what I heard... But it's still impressive af, especially since iirc the new CPU wasn't more powerful, just reprogrammable.
It’s a meta joke. Doom was about technology leading us to confront the paranormal/spiritual/other dimensional, and now the irony of AI brute forcing Doom into existence is perfect.
When I first heard about this I laughed because 'of course it would be Doom'
The technology used to get Doom to self generate inadvertently opens a portal to hell IRL....
@@ThompsonDB At the CERN large hadron collider 🤣🤣🤣 It would be PERFECT 👍🏻
@@heyimgoingtoplaysomegamescouldn't get more perfect than this indeed
@@ThompsonDB brother, observing the current worlds events, I am convinced that biblical end times are upon us... but it's not exactly what bible believers may expect... a prophecy is man-made, and man-fulfilled, by those of the same bloodline that betrayed and killed God... I do not understand their reasons, but it's not for good, and bible is not what it's supposed to be.
It's like a dream. Doomguy looked at the hall, there were armour pickups. He looked around, and the armour pickups weren't there anymore when he looked back at the hall. It happens if something catches my attention in a dream.
I swear I parked my car right here!! :P
That’s why I’m always losing stuff!
None of this is real
Fr.
Everything I've seen from ai generators feels like a dream. All of the weird "hallucinations" in pictures and conversations remind me of my disjointed illogical dreams.
The two best uses for ai-gen is uncanny-valley nightmare-fuel, and memes.
7:41 I like how the 4 armor bonuses changed into a dead zombieman
and again back to armor bonuses
Well spotted.
oof, that's not a good look for them lol
doomguys on psychedelics
We are getting dementia simulator at this point
I'm still more interested in AI run NPCs and dense populations.
It's jarring seeing "ghost towns" with NPCs in a loop, games are becoming more "realistic" but that stuff is still immersion breaking.
This. I am tired of AI generated visual garbage instead of solving real problems.
This kind of thing is an important part of that. Being able to generate new content randomized on the fly will massively increase the lived feeling of games.
@@Thedarkbunnyrabbit true, but only if the game design is good to begin with.
Hell ya, and NPC's leveling themselves up and creating their own agendas, and time in the game world still moves on and NPC's evolve even when your not playing the game..
@@lukefortune8314 Well sure, yeah. The gimmick will only carry a bad game so far.
I remember when Wolfenstein came out, then Doom, and Doom II. Man, they were so unbelievably ground breaking at the time. Doom II was the absolute GOAT.. so epic.
I would definitely play that again. Same graphics. Same 3D sound.
Ultima Underworld was released in 1992 and had much of the same technology as DOOM.
Track down Masters of Doom (audio book version is excellent) as it's the story of Carmack and Romero from commander keen through to Quake and how it all started. It's more wild and crazy than you would expect - there are thefts, fast cars, other big studios from the day make cameos it's so good. Kinda surprised it hasn't been turned into a movie or tv mini series yet.
For me it was the Quake "test" release. True 3D environment. You could look straight up, straight down with no distortions. Blew my mind. Ran great on a Pentium 90.
When it came out the graphics looked sick af. All those games😂😂looking at them now it looks like 10 pixels.
Title: "AI Doom Changes Games Forever"
Video: "AI Doom DOESN'T change games forever"
Yeah that's a quick do not recommend this channel for me
my exact thoughts.
he really didnt say anything of substance either... it's all just word vomit
@@ganondalf8090 script was probably written by ai too
You guys don't think much, do you?
Anyway, let's return to this comment thread in ten years or so, K?
@@juhalampola1954 not how games work. ai will never become a game engine.
Love to use more electricity than a small nation to run a blurry, weird version of a 31 year old DOS game
Haha- True enough. That said, I think the new TPUs are a LOT more energy efficient. The previous models were running at 200w/hr, while the new Edge ones are at 2w/hr.
I don't know if this was run on the newer TPUs, but on the plus side, it looks like we'll be seeing a pretty big reduction in power output as the newer generation rolls out. Believe me, it is something that I think about as well.
Energy is the elephant in the room
So basically it's the EV or Tesla of the video game world!?!!!
💪I create better animations on my channel than any of this AI generated JUNK!
I still have my 3.5" floppies of Doom stuffed in a box somewhere.
I'm sure your wife's box is very proud.
I bet they don't work anymore. Granted I haven't used floppy drives of that kind since probably the 90s or POSSIBLY early 2ks, but I seem to recall that if you left em sitting for long enough they wouldn't be readable when you went to use em.
Aw man, I lost all mine a long time ago :(
they are probably dead already... it's a magnetic storage and it "fades" in time...
That's awesome and sad at the same time.
I imagine, like procedural generation, this kind of tech will be handy at "filling out a world", but game developers will only end up using it in a piecemeal way cuz investors love their buzzwords, but in the mean time, you'll have curated "story zones", and then fairly random meaningless Ai-gen zones.
Oh, I for sure see that. But I can also see the modding community picking up on this. That's one of those things I always wished got bigger for a game like Skyrim-- kind of like old paper and pen RPG modules.
That might be kinda cool.
I wonder if AI could be used for the RNG of the game engine too
Sooooo we witnessing beginning proof of concept that we living on a simulation? Lol
I don't think it will be long at all before AI can generate narrative content and interesting characters to fill out those meaningless zones.
I think you will all be quite surprised at just how capable AI is becoming at making things meaningful.
DOOM was the greatest game of all time... until DOOM 2 came out and then DOOM II was the greatest game of all time and I haven't seen anything since that's dropped as many jaws
I still can't watch the movies 'Aliens' or 'Evil Dead II' without immediately remembering all the mods in DOOM - getting my hands on that MOD was one of the greatest times...
"Let's Rock!"
The new content was so good that it compensated for the shitty maps
Civilization appeared at that time too, and the original Championship Manager.
That would be true of wofenstien didn't exist.
I assume you don't know many games lol
it's really freaky to me how the AI seems to add tiny little details that are just SLIGHTLY off, but not enough to be noticeable unless you're scrutinizing. things like, the acid being a little wiggly, or bullet holes phasing on and off of walls, or the two points where walls meet being a little messy. also, the way items and barrels seem to appear and disappear when not in view.
cool stuff, but man there's something so unsettling to me about seeing AI make mistakes like this. the early footage where the monsters would melt behind the wall made my skin crawl.
I think it freaks me out because it feels like, at any second, the entire simulation could fall apart and suddenly you're being shown these horrific AI-generated amalgamations.
sounds like something that could be used as a feature in an AI generated horror game.
AI is like getting winked at by pinhead
AI has gotten as far as "DOOM, but you're having a stroke".
Its garbage, same with AI video. All shit.
Haha! Perfect description. Thanks for the laugh, brother!
I think you're being quite generous!
so basically , they are running doom , but with extra steps ?
Haha. Well…when you put it that way…
The paper keeps using the term "simulation" so it's more generating a video of the experience of playing Doom. A little like the shadows in Plato's cave.
the future of this way of rendering games depends on how efficient, accurate and demanding it will be.
@@SimplCup well if it currently generate at 20 fps on a modern computer images of a game that was made to run in 320x200 you can guess it is not efficient and will not be.
@@magen6233 well obviously it's not yet, cus they just experimented lol. They used really inefficient sd 1.4 and somehow made that architecture generate 20 fps, while there are already architectures that are way faster and higher quality than sd 1.4 like sdxl turbo for example. With sdxl turbo you can generate images of higher quality really fast probably around 5 per second if you just use the model, but I'm sure if they made slow and old 1.4 generate 20 fps they can easily make sdxl turbo generate like 300 fps or even more. The progress of ai in this 3 years is just insane, so i can't even imagine what ml models and architectures people will be able to make in 20-30 years.
If you think about it, you're essentially playing a dream or nightmare(since it's doom), where the spatial surroundings change with no continuity.
Btw. I love the breakdown of the technical stuff. Keep it up.
Will do! And for sure check out that My House.WAD video-- it's NUTS.
@@TheoreticallyMedia I think AI is perfect for Horror the unsettling pictures of people that aren't real but look real is a good example of that,
In the DOOM storyline, the UAC research facilities slowly get transformed by hellish influences as you progress. This could be very effective showing that effect in real-time.
@4thlord51 Actually, it's even better than that, it has like 90% continuity, making the changes unsettling rather than just destroying any and all sense of immersion.
@ 8:00 I would absolutely love a doom where you hallucinate level design.
Like turn around after entering a room and a staircase appears going down
dream scape for sure. Try to escape an AI film
So it's like having advanced dementia, where anything and everything can be forgotten by the time you turn around?
Yes, and it's not going to get ALL that much better than this either.
And the worst part is, before it gets any better.. we're heading for a cliff.
I played it on the UW Madison server in 1993. It was revolutionary and the first few levels were free.
I already hate this project but I want to be really clear, the RL agent and I assume also some of the visual portions of this are trained on a purpose-built fork of ZDoom, which is *not* technically original Doom. that's why there's randomly appearing/disappearing bulletholes and smudgy sprites instead of crisp pixels, there's a texture filter that is defaulted to "on" and I assume Vizidoom did not turn this off for its training model.
also found it really interesting you didn't mention the human identification portion of the paper :)
Lol whatever matrix they’re hoping to build with this tech that can count me out of….
I don't see the problem with this, it's essentially an AI powered laggy doom emulator
@@EldenLord-wh4ov Not even an emulator. It is more like a drive through google maps with occasional opportunity to feed a cat or something (which would become unfed again once you turn away, because no state is being tracked).
Btw it's not I.D. - it's id, from Freud's works
id as in kid
Agreed, and it also means 'ideas from the deep'
I never knew this. Spent 20 years getting irritated with people getting it wrong then kept hearing people say it right.
Then people like you confirming!
This is why I think that large language models are doomed, because it has to learn from humans and there are humans like you.
@@l30n.marin3r0 john carmack and everyone at id called it id like as in kid in all the video footage of them. no one from that company has ever pronounced it different. As others have said it stands for ideas from the deep. They decided it was too long so called it id. and they themselves said they likened it to Freuds' id because that is the pleasure center. So the ai probably should learn from humans like him
I think this is obviously a development dead-end, but it's neat that they got something like this working. I remember a long time thinking about a theoretical game that was produced with trillions or more pre-rendered or painted images. Like if you could paint every possible combination of pixels in the game and then the engine would just serve the correct image then GPU hardware wouldn't be a limitation and you could have graphics that are as advanced as real life, and basically be photo quality. This is kind of that idea but using ai to generate the images instead of having to make them all in advance and store them which is obviously impossible, besides that the logic of knowing what image to serve would also be impossible.
I doubt this type of technology will ever be used in the lifetime of anyone alive today unless there are totally new technologies that develop over the next 50 years that make this practical, but there is no reason at the moment to think that will happen. The limitations of todays LLM technology mean that no matter how advanced they get, no matter how fast they are, this will never work because they are just way too unreliable. Even if they somehow got 100x more reliable which is probably not going to happen, they would still be far too unreliable. Same problem that makes quantum computing probably useless for most tasks. You need literally 100% reliability from hardware, or the core of the software, for most things to work. 99.99% isn't good enough and AI is at like 30-50% accuracy for lots of things. Going from 99% accuracy to 99.9% accuracy is a 100x improvement, same as going from 99.9% to 99.99%, so as it gets more accurate it will become exponentially more difficult to make noticeable improvements.
Modern games are many orders of magnitude more complicated than DOOM, they had to train it on images from the already completed game, the hardware is something that people will probably never own and is tremendously inefficient. It's a neat POC, but definitely will never go anywhere. Honestly don't really even know why they would bother building this out besides to attract investment from tech illiterate investors.
A.I. trying to open portals to Hell through a video game.
I don't know why the AI makes Doomguy throw his eyebrows in such a sus manner all the time, but it's clearly a feature and improvement
It's "id", not "I.D."
No it’s definitely Eye Dee, you both are wrong.
It's I.D, Ideas from the Deep
Seems pretty fitting the first 'real' AI generated game would be Doom lol
@@SonicboominOnEm no it's not lol
No, it’s Id, as in Freuds theory regarding Id, Ego and Superego.
The tech is impressive from a technical point of view. But this can't create real games until the context is multiple orders of magnitude bigger. This "game" can't remember what enemies you killed. It can't remember how the room looks behind you. You don't have a map; anything that is not on the screen is destroyed and replaced with something new next time you check.
This limitation doesn't look like it's going away, and it's a total deal breaker. This can only be used for games where basically all relevant information is on the screen at hall times. So like, Tetris.
Would make for a great horror game though. Reminds me of a recent title called 'Layers of Fear'. The gimmick in that game is that as you look around a room, the game itself will actively change the room you're standing in.
As in, you look north, you see a wall. You look East, West, South, etc: just more walls. You're hemmed in. There's no escape. You're stuck here for all eternity.... or are you?
You look north again, and now there's a door in the wall. A creepy, ominous, decaying door with blood on the handle and strange sigils on the hinges. Do you open it?
You turn around and look south again, examining your options before you make a descision. Now, there's an enormous mass of writhing flesh and flailing limbs; wailing, pleading, lurching toward you. You know the game will likley de-spawn the door in the next 3 seconds. Do you go through that door?
Won’t even work for Tetris because you need a storage array to store the score…
Basically anytime you need to store a memory of anything it’ll get corrupted so this thing isn’t making game dev obsolete for a long time…
Not to mention, it’ll just put all western devs outta work but japan won’t use it…
Good thing iam moving there…
This is more of a proof-of-concept to demonstrate that they can conjure the game state in real-time. To make a complete game, it's either going to have to keep a running model of the entire game world, or have it render out finished .WAD files.
So, less like playing Doom, and more like experiencing the dreams you have after you’ve played too much Doom before going to bed (yes, I’ve done that. Still have the occasional video game dream, though I think mine look more like Unreal than Doom).
@@jasonlescalleet5611 good old unreal tournament 😊. Brings back good times
If any AI can simulate that kind of hectic gameplay I'll be very impressed. There's always so much going on on screen at anyone time - the AI would have to actually understand why bunny hoping works, the ballistic trajectory of a rocket fired off screen, the fact that health packs exist and respawn on a timer, etc, etc
We'll know humanity's truly doomed when it can simulate all that without frying the servers 😋
0:42 because of the saying *anything* can run Doom
Wow! It can, while utilizing some of the most expensive and cutting edge technology in the world, barely hit 20FPS while poorly replicating a (by modern standards of technology) extremely simplistic and primitive game that is over 30 years old! This cHaNgEs eVeRytHiNg!
It's pretty cool that they managed this, but as with all of the Generative models atm, you cant really use this to make anything new. And even more limited with this, in that all it can really do is copy a game 1:1 at best, but will most likely just introduce bugs and artifacting.
Very true, but it’s a fairly big step. Right now; we’re looking at kinda janky mods for games. But down the line I can see having a RL Agent learn a game style, and then be able to generate a different game in another style.
Granted, you won’t be making anything new- but neither has Ubisoft! Zing!
Generative models CAN generate something NEW.. Create something NEW from what already exists.. the only source of new stuff in this universe is FROM CHANGING STUFF THAT ALREADY EXISTS..
Sure, but you realize this could never power a commercial game right? What, is every single player going to pay $1-$3 PER HOUR to run this on a TPU?
"Never".
@@sw97058 moore's law it get cheaper as time goes by and newer version of that paper will come.
Hold on a second... When Doom released, it was running at 24FPS. That was vanilla Doom's sweet spot. So this thing is doing it at 4 frames a second slower than when it was brand new. Considering the radical departure from your standard 486DX2@66MHz PC Doom machine from 199X that's very, very impressive.
I think Doom was 35fps
Dooms capped at 35 fps
nope, native 35 FPS.
It was 35.
486 66 was pretty high end at the time. Most people still had 386s. It’s not like the TH-cam era of today where everyone had the latest and greatest of everything
Congratulations... YOU'VE REINVENTED SLIGE
and an even worse version to boot!
What is Slige?
@@SalgadoMaffini its an old map generator for doom
@@SalgadoMaffinisoftware that automatically generated levels
Cue the "Clowns Farting in the Basement" song. 😁😆
🎯 Key points for quick navigation:
00:00:00 *📜 The history and impact of original Doom*
- Introduction to the original Doom released in 1993,
- Explanation of how Doom was groundbreaking with its 2.5D engine,
- Discussion on Doom's continued influence through WAD files and modding.
00:02:59 *🔍 Introduction to the AI-generated game engine*
- Explanation of the first AI-powered game engine developed by Google DeepMind,
- Differentiation between usual AI-generated videos and playable AI-generation.
00:04:35 *🤖 Technical composition of the AI game engine*
- The game engine combines an RL agent and a diffusion model for AI gameplay,
- Reinforced learning allows AI to understand and play Doom.
00:06:10 *🌍 Understanding World Models in AI*
- World Models enable precise understanding and anticipation of game events,
- Distinction between AI-generated environments and procedural generation.
00:07:44 *✨ Challenges with AI gameplay consistency*
- Limitations of the AI to maintain visual and gameplay consistency,
- Degradation issues during human interaction not noted with the AI agent.
00:09:11 *🎮 AI Doom's potential and future in gaming*
- AI's impact on traditional game development and possible new branches,
- Current state limitations and future possibilities with scalable AI game engines.
00:10:21 *🎥 Why Doom was chosen for this AI research*
- Doom's technological legacy in 3D graphics makes it ideal for AI research,
- Its rich historical impact and complexity offer a challenging AI test bed.
Made with HARPA AI
I've grown fairly skeptical about the future of this type of AI. The usual line is "imagine how it will look a few years from now," but big limitations are becoming apparent both in the methodology and in the availability and quality of training data. It seems clear to me an upper limit exists for this technology and we may be already bumping against it.
I don't think it'll pass us by as a fad, but it also won't take over every field like all the hype keeps insisting.
A game that came out around the same time was Mechwarrior 2. It's a 3D simulator but had awesome graphics and was technically impressive for the time. I always kinda lump that in with games like Doom from the time
Oh, I LOVED that game! Apparently there was a bunch of rights issues that killed the franchise- which is a shame, considering how awesome that world was.
Name one person who played Doom.3 and didn't look twice at random dark corners during the.period of playing the game (and a.little bit after).
Me. But after playing Duke 3D, for a while I had the strong urge to kick open every air vent I saw and crawl inside to look for secrets.
> See ammo on the floor
> Turn away
> Turn back
> Don't see ammo on the floor anymore
Great quality
AI as usual, which means: unusable.
Dream logic.
@@pw6002 its not AI
It's a brand new technology. Do you expect it to be perfect??
@@guromenst4416 It's fundamentally broken, it's always a guesstimate of what should be the result
Cool tech demo but seems to he reinventing a worse wheel. It makes more sense to me to use ai as post process effect. Traditionally render a basic structure of topologies and scene elements. And have ai convert each frame into style of your choice. Which there has also been some research demos done using that idea.
This is more like a new type of game engine than a replacement of current engines.
It opens the door way more than that. As to say, you more than likely will still use current engines and game design, but allow the AI to generate all sorts of effects either post or on demand. For example, a real frame of the game when you move the mouse/keyboard and the real game engine is taking way too long. Instead of tring to creating a frame base off visual elements, it could be base off the game/engine it self allowing for lower input latency. Another is the fact you could add in features that take less over all elemental power from hardware such as RT that could be part of the post process.
Maybe I'm archaic in my thinking about gmaing but I find this to be the apotheosis of video game design.
I love being able to talk to another fan of a game and understand that we share an experience.
This seems like the next level of procedural generation, where players won't really be playing the same game at all.
Everyone just playing the own divergent microcosm of many games. Less and less intent and game design from the creators.
On the bright side, it provides nearly infinite possibilities for content creators.
That's just an interactive video with extra steps aka a backend doom runing suggesting prompts for the video, it still hallucinates like crazy
This is pretty impressive. I remember playing DOOM on a 486 66 mhz with 4 mb of ram.
Haha. I was one of the first kids in my neighborhood to have an EGA graphics card. 16x16 colors! Our jaws dropped!
@@TheoreticallyMedia Ah EGA, EVGA, VESA, ISA those were the days.
Why play a stable version of doom using 4mb ram when you can play a severely unstable version using 20gb vram and burn forests!
SuperVGA was a blue dream
Love how Doom went from running on potatoes to needing massive computing project to run.
Nope. I noticed on at least two occasions in clips shown early in this video that collectable items vanish when you turn around without picking them up! That's a serious glitch in a game like Doom, especially when some of those items are ammo.
And I wish people would stop calling these machine learning algorithms AI, they are NOT AI. They are just more advanced versions of algorithms that have been around since the 80's at least. I have seen them on the Sinclair Spectrum computer back then and they keep popping up through computing history with the promise of being AI but never delivering. The only difference from back then to now is the power of computers and the amounts of storage and RAM you can throw at them to make these algorithms far more effective.
One SciFi story I read recently has the right idea. They had actual AI entities that worked with humans and an enemy that was thought to be AI but turned out to be a lesser version called VI, Virtual Intelligence. The difference being AI can actually think for itself with original thoughts and be creative but VI can only mimic intelligence. It can respond as if a real intelligence, copy what it has seen, and pull from knowledge it has been taught. But it cannot create, it cannot think for itself, it is not autonomous. This is the same as our AI today, it's not AI it's VI.
Oh, 100% with you on ML vs AI- but, I think Apple pretty much proved that no one (for some reason) is willing to call it ML.
They doubled down hard on the ML term at the start of all this, and while (to be fair) there were a lot of reasons they lagged in the AI race, I don’t think their insistence to use the correct term helped them at all.
What’s the name of that story? Sound interesting!
@@TheoreticallyMedia It's an audio book on TH-cam, called Terran Contact on the channel SciFi Stories.
Also, I think the term VI is better than ML. I think people would be more willing to refer to it as Virtual Intelligence.
Whoops, wrong story title in my other reply. The actual story is The New Species. Same TH-cam channel as I said before.
I could have edited the comment but you would get notified of the change lol
So correct me if I'm wrong, the game is generated while you play it. Then you can prompt for example a "Classy Doom" where they throw champagne bottles, salads and lobsters at you wearing tuxedos and silk dresses or a "Laundry Doom" where they use washing powder, softeners and even whole washing machines to kill you? Correct?
You don’t need ai to generate a Skyrim dungeon. Generated dungeons were in Daggerfall from 1996.
"In the future, entertainment will be randomly generated" - veggietales (i think the green one)
*humour
Omni, play The Sopranos in style of Simpsons, sort Google Doc folder by Date Modified,. Build me a shitty resume season with Bart and maybe that Wendesday girl I keep hearing about on Civic. Can Paulie have a tail? Run all audio through the DonaldDuckHD.onnx model and pitch it up by 12. Oh, and no clothes. on anybody. (pulls shades down and looks into camera)
GOOGLE cancel everything I just said and just build me a world full of nude Lorraine Braccos and Edie Falcos. Helllo? (but something goes wrong and it's just Lois Griffifth from Family Guy)
Is it too soon for people to start telling the computer games industry, "Generative AI sucks and if you base games on it we will never play them?"
Isn't My House a pk3?
Also Doom 4 isn't a remake, it's a sequel.
4:28 SHOTS FIRED!!! SHOTS FIRED!!!!!
Why is the picture @1:03 AI generated?
Came here to ask that, please someone answer this! I'm seeing this a lot lately in heaps of videos.
Looks more like a shitty AI upscale
I remember that if you played Doom on PS1 and had a link cable, you could cooperate with a friend when playing Doom.
Ripping through fossil fuels with a ludicrous amount of compute to generate each frame just to end up with a mediocre game that wasnt designed by a human artist with anything to say...
Everything about this concept is terrible
Thank you
Uncivilized software engineers developing the next illegal business.
(one of )The only sane comment here
Sounds to me like we need better energy alternatives.
@@Claritism or just let artists make art?
If by "change" you mean "chalk up one more thing that used to be good until a flood of execrable AI garbage tainted even what was left of my fond memories" then yes, it does Change Games Forever.
2070:
We have successfully recreated Doom on Terminus colony, on Mars.
We can now enjoy Doom in reality.
but what is the output? As fact, doom map generators is exist since ~2014-2015, but they they were not popular at all. Because maps is fun to make.
There is another biggest problem - character animations. AI can't make high-quality sprite animation of low-resolution characters like doom or classic fallout, and this is the main problem in indie game development, where mocap is not available. And obviously no one will deal with this problem, AI will always be designed for ultra HD... So, nothing changes - map generators for uploading textures have existed for a long time already
I wonder how easy it would be to swap out each frame with a model trained on realistic images. Imagine playing the game with movie-quality frames!
"Boss, do we really need to summon the forces of hell to get material for training our model?"
@@Mewseeker"don't question me unless you want to be the sacrifice employee!"
no
that's one place it's headed.
It's impressive tech... I just don't see a point to it.
I much rather they worked on an AI that can generate proper enemy behaviours instead of wobbly bullshit graphical elements...
For sure-- but I think research like this are all the little steps involved in eventually getting better at the really good AI stuff that will be useful to game devs.
"Computer, create an adversary capable of defeating Data."
I dont think this game world generation will go anywhere, but i excpect developers will stop focusing on graphics and only one gameplay, level design and animation, essentially. The AI would then redraw the game visuals based on crude underlying graphics. Essentially the game engine would produce a controlnet and the ai could produce varying visual styles based on that. But AI image generation would need to become a lot faster and cheaper before that.
Seems like it’d be a cool way to visualize something like D&D
I understand now.
So it's soul destroying.
"Hey, what skill set can we replace with AI this month?.. I know.. game making... if...if we could let anyone create a game then everyone will want everyone else to play THEIR game and then nobody will play games anymore because they're all bored of all the games... wow.. we can do that and have them all make AI textures, AI 3D models with AI voices and AI sound FX.... so... we've destroyed the point of human creativity and uniqueness and eventually AI will have nothing new to give and people will have become so expectant of instant results that the idea of creating anything on their own will become a futile chore because someone or some machine has already done it..."
We will make AI that is more creative than humans! (Seriously, most humans have shit creativity and no interest in such things)
I personally believe Sturgeon Law will be in full effect : 90% of garbage and 10% of mad lads pushing the tech to its limit to create great stuff.
@@Mewseeker Agreed. It's the movie "Idiocracy" in the making where all the intelligent and creative people with modest intention and morals die out as the dumb spew out kids and AI garbage and become heavily reliant on a single mega corporation that bought up everything and decides what people eat, drink and love.
Ultimately a few selfish biillionaires with no concept of life (like NVIDIAs boss) caring for nothing other than bragging their own genius whilst claiming its for the world's benefit!
Grrr...
Now we only need ppl start making "great games" which get popular and not DEI virtue signaling crap which flops in the first few months. Same situation as with art.
"if we could let anyone create a game then everyone will want everyone else to play their game"
You're grossly underestimating human laziness. For instance, everyone online has access to the culmination of human knowledge, literally at their fingertips, but people are now dumber than at any point in modern history.
_Ironically, psychologists have noted a steady decline in our collective intelligence since the birth of the Internet._
In short, the vast majority of people will lack the initiative to make full use of the tech.
When AI gets to that level, most people will just make memes, and sh*tpost with it.
Only a small percentage will use it to its full potential.
Nearly shit my pants when I first saw the paper. This is actually kind of a massive breakthrough, I’m genuinely curious to see how well this tech scales
-person with no understanding of game development
@@MittzysI believe they meant it as a breakthrough in AI development, not game development, although it could lead to breakthroughs in the gaming industry. It would be like an overpowered game master, making the game as you play it in real time, adapting to your decisions. I think it would be especially fun if played with other people, like D&D. A human DM could also be present to further influence the AI. This could open a lot of possibilities for individual players. I think it would just be another way to have fun. People would still make non-AI games.
@@Mittzys -person with no understanding of machine learning
@@shizumi5243
« It would be like an overpowered gamemaster… »
Yeah, that’s basically what a human game master is and does.
Why wanting to put him/her out of the equation and get replaced by an AI ?
@@Mittzys brother I work in game development. I have a B.S. in Computer Science, and in my professional opinion, this is actually a really massive breakthrough. There are tons of cool ways that *actual programmers* could actually use this technology (granted, in a few years at least) that would make video games better. Lots of developers that I know are pretty hype about this project, and it’s kinda disappointing that instead of trying to have a conversation about it, you immediately decided to insult someone’s credibility with no basis.
Technological breakthrough in 2024: needing all of 2024s computing power to recreate an old game at 1/2 of it's 1993s framerate.
except the ai can imagine infinite levels, its not rendering the levels in a way
In real time, without human interference.
Not 100%, as to say, it opens the door for different way we can make hardware to run whatever features we want to display in a game. From everything like RT to the game it self using a more dense design than today's hardware or for at least on demand frames.
Akshually, Doom was capped at 35fps so this is 57% the framerate. /s
@@gabrielandy9272 More or less... is not an AI... which means that right now, is "imagining" stuff due to the training it had with the actual doom. This means that at some point, even with training of all games that exists, when you extend it powers to generate random levels with the same pattern, the series would basically converge.
You can argue that "well, but all levels would be 'different' even if they are the same from a high level perspective", we could assume that that might happen (they will clearly repeat but it's mathematically irrelevant how often that will happen or even if we can see it in a lifetime of playing). But then, that means that we already have that and it's called Daggerfall. Infinite dungeon crawler and questing experience! All of them are different!
And if you are thinking of that type of infinite we can totally code that already and probably make it run on a toaster for doom. Given that the toaster has a display... or... maybe it could print the frame in the toast...
So, as for right now, this is just a step up of the technology that already exists that generates videos out of a prompt. But has the potential of being interactive and assuming the demo is not "slightly edited" like we've seen in the past with some models, much faster since it goes in real time and when, could go forever.
This is very fascinating. I'll probably just go ahead and mark my own words that I don't think this is really going to go anywhere as far as companies selling GPUs that will generate instead of render. If it ends up that way and we get insane visuals with low energy cost and consistent framer rates then I'm all for it. As it should be noted, though, all of these things are well outside of my knowledge ability.
Absolutely love the idea of "Can it run DOOM?" being the go-to benchmark for every new piece of software out there
2:43 doom 2016 is NOT a remake, its a reboot its an entire original game and was originally called doom 4
Its not a reboot. It's a sequel to Doom 64
Its the same marine
@@christopherpekel6096nah, its a reboot. Doom 2016 is the updated experience of playing doom 2 1000 times and wanting more challenges.
Doom 3 is the actual remake of doom trying to put more realism and effects into the survival/horror action game. There's more fans who want action over horror it seems...
Is not a reboot nor a remake is a continuation
Doom 2016 is a sequel, Doomguy is resurrected/reawakened in the beginning.
@@gesi7072 a reboot is when the original concept is reimagined, with core elements retained. Slayer is a new character, the lore is reimagined etc.
To be an actual sequel would require the same gameplay, characters etc and overall setting of the originals.
Doom 3 is also a reboot of the original. I don't think people understand what reboot means. They can still link slayer to the older games, but its a different continuity.
AI is regurgitation and we are being told it's fresh food. The game itself and training on that game are still necessary.
nope, why do we need low fps, hallucinating DOOM while we already have Brutal DOOM with HD textures and dynamic lighting running 60+ fps on high resolution? Admit it already, large models does not work for what you are trying to use it.
The emperor has no clothes.
I mean that AI didn't make doom from scratch, it needed the existing massive data set of doom to train on. It super cool but in its current form it doesn't mean much imo until it can make games from a collection of multiple games and mixing the styles and genres.
Counter Point:
Companies/Publishers like Ubisoft will /are absolutely be training AI to generate entire "open" worlds based on their own previous IP, and eventually even the points of interest/quests/missions.
Specifically, if for nothing else but to significantly reduce the cost of level development.
@@G4M5T3R No publishers they are using ai make the entire game, they use it to help generate specific assets, not the entire game, big difference. They are not wasting money on something like Nvida did here, a game engine for a game that is 30 years old gets worse the more you play it.... Nvidia has cash to burn on research these development studios/publishers, don't. This toy project had to cost a few mill with zero return.. Fundamentally hallucinations are baked into generative AI with the current approach. Untli they can solve the hallucinations these problems will persist. Unreal's gen AI solutions are far more practical and valuable than this research demo. Generating an entire game on the fly then testing it, looking for flaws, nightmare. Generating an asset and then looking for flaws great solution.
@@G4M5T3R Generating an open world is not generating a game. They can already generate worlds, without the massive cost of models. No mans sky for example. I think there is confusion around the word generate, the game is being generated on the fly by the Nvidia model vs generating a world during development. Very different things. Nvidia example here is generating the next frame in real-time. Its a novel experiment.
@@G4M5T3R I guarantee you that any dev shop could build doom cheaper than the process used here and with better results. You are assuming it was somehow cheap.
I still have a hard time thinking about ANY of this as actually AI.
Something had to brute force learn how to play doom to figure out what was allowed and what was not.
Something had to copy the graphics & characters from the original game. It had to be cajoled into behaving with the stats at bottom of screen.
And then it what...it just reproduces the world & elements that it's copied from the original, based on inputs and actions from the player?
Is this not just some fancy way of using machine learning to copy an existing game the long way round?
To be fair, it's not even copying the game so it fails even at that. It just imitates it. What you get is frames, but no enemies, textures, doors, places, variables, states, points, etc are stored anywhere.
Imagine AI accurately redraw and enhancing all your ps1 games with photorealistic graphics🤯
@@NostalgicPlayYT imagine the crash remake lol :
1)can't hallucinate
2)no blurry wobbly images
3)runs locally on a toaster (compared to the supercomputer needed here)
4)the only training needed is for the player.
5) relatively low effort , hig results .
6) fine tuning and control on the overall result.
7) don't require a roket scientist to play it.
Imagine the AI Sony lawyers on your doorstep
That's probably easier than this Doom thing tbh.
How about RPGMaker games? How about the original Pokémon?
@@Zepoz imagine a lawyer trying to sue a robot... makes good sense🧐
So basically, they’ve introduced Skynet to Doom.
Great.
I don't think AI can ever make a proper game. There are too many smaller details from a game design perspective (jump height to platform design, secrets, how do you get out of a stuck area, etc) that AI can't properly solve in a cohesive manner. I think the concept is interesting, but a human mind is still needed to thoroughly fix all of those flaws
For sure. I think that an interesting future will be when anyone can create the game they want to play- but I guess the thing that isn’t in the sales material is that everyone will also suddenly become QC!
I think that if you were unable to accurately envision the 'impossibilities' which have become not only possible, but almost-mundane in the past year alone-i.e., your name doesn't end with "Karpathy" or "Sutskever"-then to impose that same normalcy-bias on future potential constitutes a failure of introspection.
If you wish to remedy this metacognitive blind-spot that seemingly 98% of humanity, and certainly most TH-cam commenters apparently share (which, assuming you value your sanity, I strongly recommend doing), I would suggest Eliezer Yudkowsky's seminal essay "There Is No Fire Alarm for AGI".
I liked Nvidia's approach to this. They basically use a traditional game, gta 5 in this study, as the basis and then use AI to generate the final graphics.
In theory this would allow developers in the future to make a game with very basic models and no textures and than just assign prompts to these simple models like "car" or "tree"
You can then play this game carefully crafted by the devs while the graphics just get generated on the fly.
The only thing an AI will never be able to do is really feel emotions or reach sentience. everything else an AI will eventually be able to do, because what humans do isn't that impressive.
Image generation barely existed in 2021 and here you go saying "AI will never do this" in 2024. See you in 2025 i guess
Does that mean in future, Some games will have options to run on Tenser Cores or CUDA cores completely?
As far as i know, CUDA is faster than the GPU itself, so how far this can go in the future will be astonishing!
I don't think there's even an ounce of practicality to this POC. This is just a "is this possible?" project, and nothing more. The hardware one is going to need to have at home to run the most recent games of the era is ridiculous. So unless there's a *massive* technological breakthrough that suddenly makes quantum computers affordable to the average Joe, I don't see this concept beating game engines, ever.
something unsettling about training an ai on doom
Haha, there is that as well. Hey, at least it wasn’t the Terminator game!
We jumped from shotty ai generated portraits on steam to whole ai generated games. Human art is burning, and we are fueling the fire
A mod for Skyrim adds an AI companion. That NPC's dialog is generated in real-time, depending upon how the player responds, asks non-scripted questions, etc. Another AI mod allows non-companion NPCs to respond to the player in real-time as well. Yes, real-time, AI gaming is coming
Real time but there's always delays 😅
I really hope not. If using ai slop in games will become a "norm (the soulless corporate gobshites' success at destroying another subgenre of HUMAN art. Since only real HUMAN beings with thoughts and feeling can and always WILL create ART)" in the future, I just won't be playing future games, lmao.
@@croc_moat2327 > I just won't be playing future games, lmao.
Cool story bro
@@croc_moat2327 Then don't play them and leave the other people having fun alone
This vid makes me appreciate that I don't see many ai vids
Let’s use a thousand servers consuming a gigawatt to produce the same result that a single 386SX could turn out in 1993. AI is so amazing! Stock price up and to the right!
"How can we waste as much cpu power as humanly possible to do simple things a calculator can do?
The statement "games in the future will be generated, not rendered," is nothing more than a tech-bro catchphrase for people who have no idea what making video games is like.
No, once ai and hardware is advanced enough games will be randomly generated or generated based on the users prompt on the fly.
With that being said that doesn’t mean conventional game development won’t exist, but ai will still play a massive role in that as well, the endgame is eventually ai just making its own games in a matter of seconds that would take an entire team of people years to do.
@@DoritosBurger More that there will be an established engine, assets, etc. for the game that an AI will then use to randomly generate things. Because a game trying to simulate everything from the ground up is just asking to crash randomly at the tiniest of problems cropping up. AI isn't and can't be perfect because it's not a thinking entity that can self evaluate to find errors. That's why even when you're asking something like Chat GPT while it can give you a textbook answer, because it's copying a bunch of textbooks, if at any point there's a deviation anywhere it can quickly break down into gibberish. Just look at all the silly videos on youtube where AI bots on stuff like Google give wrong our outright destructive answers, now apply that to a game where something destructive could cause a game to crash, saves to corrupt, or even blue screening a computer.
Until AI can self evaluate to find errors and adjust itself that's always going to be an issuee. The issue with self adjusting AI is that's a hop and a skip away from self determining AI, which is much, much bigger problem.
@@ComotoseOnAnime you’re thinking of todays ai though, I’m talking about ai that will exist in 50 to 100 years.
@@DoritosBurger And like I said, an AI will *always* have this problem unless they are smart enough to fix the problem. If an AI is smart enough to fix it's own code it's probably smart enough to go beyond the bounds of it's own code, and then we get into the proverbial paperclip AI territory unless protections are in place.
Further, an AI generating an entire games engine and assets on the fly is incredibly resource intensive and wasteful. Even 50 to 100 years from now that kind of generation is going to be resource intensive by default. Having the generative AI built within the engine rather than the engine being generated by the AI makes the systems run by the AI an order of magnitude less resource intensive. It would also let them have a memory longer than the 3 seconds in this video without ballooning out into ridiculous storage requirements.
Just look at procedurally generated games like minecraft, it's not all that resource intensive and the world isn't all that hyper detailed, but I've seen world saves that can get up to 2-3 GB in size. Now expand that out into a more high fidelity game with persistent assets like wholly unique towns and NPC's where the game is keeping track of that information and you could feasibly have a game save that gets into the hundreds of GB to even several TB's of space, for a single save. Unless that 4 dimensional crystal storage tech I heard about the other day becomes viable for home use in the next couple decades our storage capability is starting to slow down, it's not making jumps in leaps and bounds but in little steps here and there, and those kinds of file sizes, never mind the power and cooling requirements to run those kinds of systems starts to become infeasible.
Yeah. You need training data
boycott all video games with generative AI content
As a gamer and game programmer, I believe this is not necessary because experts need to put more efforts to other aspects. Procedural generation is great, and still can be made greater. In same time, gamers need better AI as opponent, or for cooperation. What I mean better is not something like god-like AI, but human-like AI. Funny most games now have great graphics, physics, etc but cant have great AI.
People need to realize that AI isn't smart but incredibly dumb, as is also expressed by the huge computing effort needed to achieve results that then human intelligence has to repair. We wouldn't feel the need for AI if humankind wasn't dumbed down in its abilities to shape smart algorithms, because what we mostly do is offload neural networking in a brute force style.
I do believe we've found the *least* efficient way to run DOOM.
There are hundreds of trash games released each day, can't wait til it's tens of thousands of trash games made by AI.
So a really crappy and slow version of classic DOOM, where there's no intentional map design and your ammo counters are flickering between random numbers. Sounds more like a DOOM-inspired nightmare to me than an actual game.
@@fat6776
Let tech bros get hyped, their lack of talent and imagination don’t let them get hyped so often.
@@pw6002 Lack of talent? What you on about? Don't you know it's a law that Doom has to be ported to everything? It's true, it's mankind's duty to put Doom on everything with a screen. Just check the UN charter, it's there, trust me bro.
@@the-answer-is-42
I was talking about the lack of talent that is inherent to AI tech bros, not about the fact that DOOM indeed is ported on every existing electronic device in existence ! ;)
Every proof of concept/ tech demo is crappy and slow at first. This is supposed to show what the potential is not something that's ready to use
@@pw6002 And I was making a joke because I was in a silly mood. I agree some of this AI stuff lack talent, though, but I'll let AI Doom slide for the memes.
6:58 the beginning stairs with the armor transform into the beginning of the game. That's trippy.
what a way to overcomplicate things unnecessarily.
it takes exponentially more calculations and provided information to render something *_to specification_* starting from a blank slate, than it takes from a pre-defined model.
which is exactly how Doom was designed: "Render [ceiling] from [layout] and [information about height]"
if you want the engine to render something it has no idea how to render, you have to TEACH it (AKA "program" it!) how to render it! which you'd be doing WHEN you actually program it!
if you MUST correct the engine to render what you WANT to render, accurately and to specification, might as well write it once, set in stone, and direct the engine to IT every time.
WHICH IS WHAT IS DONE ACROSS THE BOARD, BECAUSE IT *IS* THE MOST EFFICIENT WAY TO DO IT!
Exactly. A lazy programmer is a good programmer because they learn to optimize everything to save themselves the extra effort and headaches they'd have in the future. This application of AI is a deadend.
Not seeing it happening. And I hope it doesn't.I mean, why using a defective by default technology for an application that is much better solved with traditional approaches?
wad is map upside-down
What a waste of resources.
you're missing the point.
first of all, this is only the first generation of the engine.
2nd, imagine your MOST favorite game ever that will NEVER end.
3rd, think how easy it will be to create (or generate) a whole new AAA game designed exactly like you told it, and now you can sell it and make money.
EVERYBODY will be able to create full Triple A games with the highest qualities and the best stories.
WOW you made a game that takes next to no resources to play and made it take 10,000,000% more processing power, good job 😆
For real though, of course this is technically super cool and impressive and I'm very excited for the future of GameNGen!
It's amazing seeing this video after something that was shared with me a month ago. I've been coding for 26 years now, mostly focused in AI and networking, so I have lots of friends that specialize in those areas. Last month a few friends of mine showed me this project they've been working on. They started by showing me this FPS with characters made up of about the same geometric density of a Minecraft character and the backgrounds was pretty much flat, untextured, polygonal primitives. So I'm like, "What is this? You guys making some kind of retro Sega Saturn FPS?" My friend fires up the generator engine and suddenly the rendering looks like some kind of hyper-realistic film from a first person perspective. It was running around 5-10 fps, but it was fully functional. I was just standing there looking lost as I couldn't believe they managed to optimize all that data to the point that the game was reasonably playable.
My friend says to me, "Remember when you kind of postulated the idea that one day we'd use AI to generate content for game backgrounds instead of rendering vast distances? Yea, we're actually able to do it for everything just really slow right now." I was like, "Yea, I didn't see anything like this coming for a while though." He smirked at me and said, "Imagine where it'll be in 5-10 years." I can only imagine what gaming experiences will be like for my kids in the coming decades.
Nope, this is bs. Plus it was done by Google so most def this was faked for a presentation.
Longer version - Im a game developer (20+ years) and also work with AI. AI in it's current form is a glorified autocompleter. In order for AI to produce a game like in the video, you would have to feed it (machine learning) with the original source code of Doom, which is most likely what Google did since it's open source now). If you feed it with the source code, producing Doom levels is trivial - it also loses the impact of what this video is trying to convey "making doom with prompts". RIght, feed AI with the source code of any game and spend countless hours "teaching the AI to autocomplete stuff" and then for sure you can generate levels of any game - that you fed the source code into.
TDLR the video is bs, coming from Google, not the youtuber.
I was just reading the Google paper on this - still sounds like BS to me from the same company spilling bs about their non existing quantum computer. But oh well :-)
@@yohami Er no, this is real.
@@yohamiso what you are saying is that your original reaction was wrong after reading the actual paper. But you stick to your original opinion purely based on that it is Google? Might want to relax a bit there and maybe accept that sometimes innovation and real new things can come frome big companies. They are still filled with actual people you know.
Yeah, you have made some top games... I was developing space marines when you was thinking of a computer.... Zzzz
colossal waste of resources
What long winded unwatchable crap. Im still searching for video of gameplay. This is not the sistine chapel its a video game, stop making this so precious.
I like that the last word of the video is actually your name. Most of the people start with their name first and who they are first. I like your "approach" of introducing yourself.
I would say this is the dark ages of Doom mods where people stop creating and instead made an AI make it poorly for them.
Tho I know 90% of mod makers will stick with doing it themselves.
I wouldn't be too worried, lol. Nobody sane would use this for anything gaming-related besides maybe a completely new genre of "randomizer" type runs.
@@restudios9612 exactly my thoughts