Before reading this article, I would ask you to watch this video.
Done? Then let’s keep going.
What you’ve just watched is Google’s latest LSD trip in the world of gaming: GameNGen, or, more specifically, Doom running in GameNGen.
GameNGen is a neural network attempting to dream up a whole game with absolutely no game code, you just train the AI and you have the game.
In the case of Doom, this results in something that kinda looks like the original, except items and enemies just appear and disappear whenever they feel like, the image randomly gets smudged like some poured some Vaseline in your eyes, and it runs at “over 20 frames per second on a single TPU”.
Now, that framerate is pretty bad, but how bad is it? Well, consider this: the original DOS version of DOOM ran at 35fps on an Intel 486, a processor released in 1989. (that’s 35 years ago!)
However, let’s also take a look at what hardware this runs on. The main GameNGen website says it runs on a TPU (Tensor Processing Unit), which is the name for Google’s custom-made hardware for machine learning. In the paper they wrote about the whole thing, they specify they’re running it on a TPU-v5. While Google doesn’t sell this hardware to give us a hard price for it, it is rather comparable to Nvidia’s H100 AI cards, which cost roughly 25,000 USD.
So that’s about 25 grand just to play Doom with a worse framerate than what was achievable on a computer old enough to be married and have kids, with artifacts that would make you think someone sneaked some LSD in your drink. Gotta love technological progress!
Google claims that “Human raters are only slightly better than random chance at distinguishing short clips of the game from clips of the simulation”, but if you’ve watched the video they provided then you’re probably asking yourself the following: have said “human raters” actually seen, let alone played, any video games in their life? You really don’t need to be familiar with Doom to know that enemies, items, and projectiles are not supposed to pop in and out of existence, seemingly at random.
Watching the other clips they have on the website, there are even more obvious tells that this is Stable Difusion, like the severe drop in video quality when the player climbs up the stairs or falls from the ledge in the bellow video.
Oh, and of course, the enemy pop-in later in the video is even more jarring in this video than in the first vid.
However, Google’s GameNGen isn’t the only attempt at turning neural networks into game engines. More recently, GameGen-O released a video showcasing their AI generating open world games.
According to some sources, GameGen-O is Tencent’s doing, however, no mention of the company is made on the project’s official website (Archive.org snapshot, the original site returns a 404 at the time of writing).
One thing you’ll notice right away in the clips shown on the site is that most of them are only 4 seconds long. One could speculate this is because showing longer clips would result in the videos devolving into acid trips, but… well, there’s no ‘buts’ here, even in these short clips we can see classic AI generation artifacts like nonsensical streetlights, or Geralt seemingly walking on water.
And one can’t not notice the terrain deformation on the right side of the screen, or the foliage constantly changing its shape.
Another issue is that all of them are just regurgitating existing games, namely Cyberpunk 2077 and The Witcher 3 respectively in the above clips. In other words: textbook slop content. Generating slop games with AI might be a dream for publishers and franchises constantly re-releasing the same game under a different badge (looking at you, Call of Duty and everything from EA Sports), but most definitely won’t result in good games.
All of this, however, ties into the larger problem: AI research almost universally focuses on use cases that look flashy and seem like sci-fi tech, but end up being extremely wasteful, and either utterly useless, or a net downgrade from doing things without AI.
Now, could AI be used in video games in not utterly useless and wasteful ways? Sure! Alien: Isolation, while not powered by a neural network, did use machine learning to make the Alien seem a lot smarter by making it remember what tricks you used to distract it, where it last saw you, and how you escaped it in the past.
Player companions could also be given a lot more depth with machine learning by making them remember past adventures and bringing them up in conversation, or by learning the player’s preferred play style and adapting to best support the player (say, if the player prefers a stealthier approach, then the companion can also gravitate towards stealth).
However, that won’t generate headlines like an AI-generated Doom LSD trip, so don’t expect any of those uses to be the case when “AI” and “video gaming” are in the same sentence.