New AI Model 'Learns' How To Simulate Super Mario Bros. From Video Footage (arstechnica.com) 31
An anonymous reader quotes a report from Ars Technica: Last month, Google's GameNGen AI model showed that generalized image diffusion techniques can be used to generate a passable, playable version of Doom. Now, researchers are using some similar techniques with a model called MarioVGG to see if an AI model can generate plausible video of Super Mario Bros. in response to user inputs. The results of the MarioVGG model -- available as a pre-print paper (PDF) published by the crypto-adjacent AI company Virtuals Protocol -- still display a lot of apparent glitches, and it's too slow for anything approaching real-time gameplay at the moment. But the results show how even a limited model can infer some impressive physics and gameplay dynamics just from studying a bit of video and input data. The researchers hope this represents a first step toward "producing and demonstrating a reliable and controllable video game generator," or possibly even "replacing game development and game engines completely using video generation models" in the future.
Re: (Score:2)
Yeah, i'm sure there're interesting ways to use LLMs, generative AI etc in game implementation / development. But i don't understand how this particular approach plays into that even theoretically.
Maybe in the far future you could say "Go watch some Twitch and make me a (legally distinct) clone of Fortnite / Fruit Ninja / Animal Crossing" ? is that that what they're thinking? I mean... that probably HAS been the prompt for various legit studios/teams over the past decades ... so ... maybe. But i still
Re: (Score:2)
I think they are thinking marketing; however, there is something to trying to infer behavior from observation -- we can infer the game model in our minds and imagine it in our minds as well as variations upon that model. The physics and so on. Progress on that is worthwhile; however, it's still just improved mimicry and not serious intelligence. But what is? perhaps we're just multiple levels ahead of mere aping prior experience? We used to think Chess was some high brain level activity but it turns out i
Re: (Score:2)
I can kinda see this working if you train the model on 1 game, like doom, so people can play unlimited levels of doom, but how could AI ever use this to generate a new game that's even remotely coherent, clever, challenging?
We saw what happened to Concord this month, a game that was merely OK but not great. These AI games will be less than OK.
Even with endlessly generating an old game... Some games like Binding Of Isaac and Void Bastards could really use that, but games where every level is designed with gr
call me crazy, BUT (Score:5, Insightful)
Re: (Score:2)
they'll be entitled to no less than $2m in compensatory damages and $2m in punitive damages.
Another meaningless stunt (Score:2)
It has no practical application except to create the illusion of an AI-based product.
Re: Another meaningless stunt (Score:3, Insightful)
Re: (Score:2)
Exactly.
Re: (Score:3)
Not really. Once the model has enough parameters memorized, new games can be generated from existing games just by altering the parameter weights.
Re: (Score:2)
Hahahaha, maybe really crappy ones. Good luck with that.
Re: (Score:2)
> It has no practical application
Entertainment! Many love it when bots F up and draw/make/do silly things. It's similar to the joy Jedi's must feel when Darth Vader trips.
Your Job is on the AI list (Score:2)
first step toward "producing and demonstrating a reliable and controllable video game generator,"
In a way I was waiting for this in video game development where years and dollars can be spent only to have it go nowhere or delivered so late that it is no longer wanted. If AI can speed up development of other things that take years bring it on.
"Replacing game development and game engines" (Score:2)
Re: (Score:2)
Re: (Score:2)
Tetris or Chess absolutely could be done by this thing, the key is that all the important state information is on screen at all times. however it still won't actually understand the rules so it's a matter of chance whether each move is actually legal.
I have a table of requirements located at %var1 and demonstration video of gameplay at %var2. Review the gameplay video and requirements document now and generate a new %OS_based game in %game_generator_tool and output the result to %var3. Save detected visual assets individually in folder %var3 and append %var1 to create description and related link to each detected asset. At midnight every night after the first creation of %new_game review the requirements in %var1 and linked assets in %var3 to review
Re: "Replacing game development and game engines" (Score:1)
Tetris rules are pretty simple and could be done quickly with this technique. Tetripz on the other hand I would like to see.
Chess rules are simple but the game evolves in complex ways, you could derive legal moves very quickly but technique will just be imitation.
*yawn* (Score:1)
Re: (Score:2)
Re: (Score:2)
I'll get rid of my anti-Trump siglines when Trump is (1) defeated, and (2) imprisoned for his many crimes against this country, and (3) preferably dead. Until then, it's touch shit for everyone if you don't like it.
Re: (Score:2)
Wrong direction (Score:4, Insightful)
AI is useful for coding, I'll give it that. I use it, it's super helpful. Why don't they improve it in that direction, what its doing now will end up in an uncanny and unwanted plateau. Show a video of Super Mario and have it edit or create an existing platformer game codebase to incorporate the characters and behaviors seen in Super Mario? The reason that is superior is because it will have less glitches and is easier correctable by a human (and it can learn from that). Same reason why having AI generate movies is dumber than having it create a 3D scene file with all the objects or 2D graphic image with editable layers/vectors. The output of AI needs to be more human-editable.
Re: (Score:2)
This is an inevitable next step in the evolution of this technology. AI will soon be able to reverse engineer complete games, reproducing assets and gameplay behavior as necessary.
In other words, the results will be indistinguishable from taking an existing game and then paying a Chinese coding farm to create a copy of that game with different assets.
Re: (Score:2)
That's not what this is doing, though. This has nothing to do with the game code. They're having the AI study the video and player inputs and based on that generate a real-time(-ish) series of images that look similar to the original game and reacts to user input in a manner similar to the original game. Think of it as a series of Midjourney prompts.
Draw a frame that looks like a screenshot from Super Mario Bros.
Player pushes right
Now draw the previous frame with the viewport shifted five pixels to the
"Crypto-adjacent AI Company" (Score:2)
Can it only copy? (Score:3)
It seems they are having to train it on a completed, fully working game in order to generate a semi-functional facsimile of that game.
You could say the same thing about all the LLMs. They don't understand the logic behind the words, why would this game replicator understand the logic behind the images?
Re: (Score:2)
Nintendo to sue in? (Score:2)
Nintendo to sue in?
I saw the doom video (Score:2)
Re: (Score:2)
Porting function (Score:2)
I can see value for this when porting. As many have said, you basically have to make a game to get the video for an AI to ingest. But, with something like this I can see you feeding it video of whatever game, and it porting it to every other system available. I doubt it would be a single click effort, but if it works as described, could likely cut porting effort down by a considerable margin.