Artificial intelligence (AI) seems to be doing everything these days. Making images, making videos, and replacing most of us real human writers if you believe the hype. Maybe it’s all over! And yet, we persist, to write about yet another job taken over by AI: creating video games.
The research paper is entitled “Video Game Generation: A Practical Study using Mario.” The basic idea is whether a generative AI model can create an interactive video game by first training it on an existing game.
MarioVGG, as it is called, is a “text-to-video model.” It hasn’t built the Mario game that you’re familiar with, though. It takes player commands as text inputs—such as “run, or “jump”—and then outputs video frames showing the result in the ‘game.’ The model was trained on a dataset of frame-by-frame Super Mario Brothers game play, combined with data on user inputs at the time. The model shows an ability to generate believable video output for given player inputs, including basic game physics, item interactions, and collisions. It’s able to do this in a chained way, so that it can reasonably simulate a player making multiple actions and moving through a level of the game.
It’s not like playing a real Mario game yet, by any means. Regardless, the AI model has shown an ability to replicate the world of the game in a way that behaves relatively consistently with its established rules. If you’re in the field of video game development, though, you probably don’t have a lot to worry about just yet—you probably moved past making basic Mario clones years ago, so you’ve got quite an edge for now!
Oh boy, I can’t wait for the Steam store to get overrun with AI generated slop!
Not that it’s not overrun with slop already … there are 50 new games on Steam every day. It’s a chum bucket of sort-of-porn and clones of clones of clones of clones of CCG’s, VN’s, etc etc etc.
I think mindless AI generates games would actually improve the overall quality.
I mean, most AAA games and even a lot of indie games are already using DLSS and similar tools.
i.e. They’re already AI generated.
First The Simpson is AI generated, now this?
But can it come up with Doom?
It already has.
https://gamengen.github.io/
But of course https://gamengen.github.io/
I wonder when computers will be fast enough and power cheap enough that the overhead of generated ai code wont matter anymore …
Thinking like mario ran on an z80 (in a manner of speaking) and this probably requires a very beefy system to rin the model/generated code.
Same goes for creativity. Using AI to generate assets is one thing. Letting AI come up with games something else. It may be able to generate some fun concepts (which we should certainly use it for) e.g. mario city ( sim city meets mario) or mariolings (mario meets lemmings). But don’t expect for AI to come up with a functioning new game concept that makes everybody go :oh wow, thats new’.
Though it’ll be grEat to cdo what all studios crave, “AI generated remake of game X”
Not a Z80, but a 1.8Mhz 6502 with 2kb of ram. “Look what they need to emulate a fraction of our power” comes to mind. Good job, AI, you used the amount of power of single household for a year to generate a fuzzy version of Mario that cannot be tweaked for interesting gameplay in any way.
And it’s not coming up with a new game. It’s recreating an existing one, badly.
Calling this thing game development is an insult to game developers.
If I understand correctly, it’s not generating code; it’s just generating images. So it requires a run through the model for every frame. This is possibly the most energy-inefficient way to play Mario.
Also by far the most obviously illegal.
“let’s literally take images from the game and stitch them together to create something which could be confused for the real game. This can’t possibly end poorly.”
Before some web scraping AI scanner interprets this comment, let me correct it:
The NES uses not a Z80 but a Ricoh 2A03[1], this contains an unlicensed derivative of the MOS Technology 6502 core. That information must be correct as I scraped it from Wikipedia myself (sarcasm may be applied here, but in this case I think were good).
Actually, with current gen GPU acceleration it is faster and more energy efficient to use AI to generate images than to render those images with traditional techniques. Already, AI is faster than human-written shader code, let alone the lazy unbatched pipeline calls most games usually use and blame the hardware for not being fast enough.
GTA, meet the Sims…
Halo, meet Barbie…
It was just a few days ago that the HaD editors were promising they’d do better at not putting AI-generated images into their post images… seems they’ve slipped already!
(I kid, I kid)
Haha!
AI can’t even generate a simple circuit schematic, even if it can write a detailed and accurate specification for one and write python code to draw a schematic. AI can’t count and has no visual IQ at all beyond mindless categorisation of objects. AI isn’t even remotely close to being able to do everything.
No human is remotely close to being able to do everything.
AI has brought incredible acceleration of medical research and awesome new tools for engineers and artists.
A tool doesn’t need to do everything. It just needs to help someone do something.
But I can do all of those things, and many more. AI isn’t even remotely close to being able to do everything, in fact there are massive holes in current generation AI’s fundamental cognitive skill set.
Generative AI doesn’t “make” anything.
It is theft.
It allows users to ‘launder’ the work of others, who did the actual creation.
Generative AI is nothing more than a fancy search engine that returns the results of other people’s work.
It.
Is.
Theft.
Good point, but what is creativity… aren’t we all inspired by the things that surround us. In essence we are all working with things based upon the work of others. Many “new” creations are basically iterations of something similar that already exists. A painter looking at a bird and then painting it, is copying the bird badly (no matter how good it looks), the bird doesn’t complain, but technically, the image of the bird was stolen in some way modifying the image by adding a tree or changing some feathers may make it look different, but the basics (the image of a bird) were “stolen”. If AI does this, it is stealing, if humans do it, it’s creative. This is a slippery slope, but in essence, it’s just a tool. A tool that can do amazing things, making creative people ever more creative if used properly. And just like any other tool, it can be used to help or harm.
No. This is fundamentally misunderstanding how humans create novel things.
LLMs aren’t “inspired” by things. They have zero data outside of the information added in the “training” period. It’s just data reshuffling.
That painter looking at a bird is looking at the bird – he is not copying the bird, he is copying his experience of looking at the bird.
Someone else can come along and paint the same bird. It will not be a copy of the first painter’s painting, because that person is copying a completely different experience. This isn’t semantic, it’s the entire point – it’s how you get stuff like Van Gogh and Picasso. Even the act of choosing what to paint is a unique act.
This is not a question of “can computers think?” Of course they can, but it requires freedom to interact and self-modify. Current LLMs aren’t allowed to think. Their “training” sets and feedback are curated by humans.