-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
Grok 3 can now generate rudimentary 2d videogames
- Thread starter BennyBlanco
- Start date
Pasta la Vista
Member
Exactly, all games made like this should be free.
Jeff Spicoli
Member
At 42 i feel already old for this world.
Zannegan
Member
If this works, they won't want AI making games that you buy. Instead, they'll want you subbed to the AI itself for life, which will constantly spin up ad-filled evanescent fantasies for you, like TikTok meets Dreams. If they have their way, it will be too expensive for anyone but a handful of big players to create or maintain competing AI and also too expensive for anyone to self-fund the development traditional games. Then there's the added bonus that their audience would be so addicted to the format and have such a short attention span that long form gaming would seem as accessible and attractive as reading War and Peace...
...not that any of that will happen, per se, but it would be a megacorp's wet dream.
EDIT: The War and Peace comment may be a tad dramatic. Lol. Maybe a more realistic comparison would be a complex board game with lots of parts, rounds, and rules--incredibly unique and rewarding experience, but requiring an initial investment of time, money, and thought that will discourage most from ever experiencing it (not unlike War and Peace, actually).
Last edited:
Some games are just nevee gon a be possible without AI.
You will never have an open world game where every NPC has his own personality. But that would be possible with something like chatgpt integration already for the most part.
It would be really cool in a detective game like LA Noir but the crimes woulsnt have fixed solutions. Instead you'd do what a real life detective would do. You could interrogate anyone, and ask them anything and get a realistic answer anyone would have their own personality and their own limited knowledge, like yeah a guy that lives, works and does everything in the other side of town isnt gonna know shit about a murder that happened on this side of town.
99% of the game is still made by humans.
Last edited:
CompleteGlobalSaturation
Member
As long as I get games with as much quality as MGS1/2/3, God of War 1/2/3, Uncharted 2, The Last Of Us, Mass Effect 2/3, Doom 2016 and Doom Eternal - then I don't care what or who has made my games (AI assisted or human).
I'd be sceptical of games maintaining their quality without significant human intervention.
A lot of A.I. generated art is crap.
I would think A.I. would be more useful at rendering/textures/models with art made by actual humans.
I'd be sceptical of games maintaining their quality without significant human intervention.
A lot of A.I. generated art is crap.
I would think A.I. would be more useful at rendering/textures/models with art made by actual humans.
Last edited:
Jack Videogames
Member
Question, is this really new? I asked ChatGPT to make a game in html5 for me and (as always with chatgpt) it took several tries for it to understand the instructions but he made a small, blocky looking but playable platformer
this is an oversimplification.
not conceptually but they might have tweaked grok to handle it better
Unknown Soldier
Member
Soon AI generated games will be better than average Western AAA woke political propaganda simulator
Knightime_X
Member
You're overlooking the bigger picture and vastly underestimating what's in store.
"80's atari games
AI still in its early years...
Just replace Will Smith and spaghetti with video games.
It WILL get better and better. We're nowhere near the finish line.
Last edited:
DownLikeBCPowder
Member
Exactly right. Again, I point to Oasis - though it was specifically trained on Minecraft videos, it still does not have access to the code, but makes an amazing approximation. I'd argue it's engaging for the psychadelic nature as well, I would absolutely play "approximations" like this, but regardless the fruit is coming soon. People either attack it because they don't like it, can't understand it, both, or other reasons.
ScHlAuChi
Member
It seems more like you are completely overestimating teh tech or not understanding how it works at all!
One is a non interactive video, the other is a fully interactive predetermined world.
LLM´s CANT get better without more data, where is the new data coming from?
ScHlAuChi
Member
In order to "approximate" Minecraft..... .....Minecraft had to exist in the first place!
You dont see the problem?
People dont attack the models, they attack the stupid marketing behind it, making people believe things it can do that it cant!
Possibility Space
Last edited:
DownLikeBCPowder
Member
Nope. Sounds like the argument from people who think all art is created in a vacuum. Everything is built upon what exists prior whether that comes from a human or a machine.
Hayabusabruh
Member
How did you ghett to that conclusion? nvm. I tried.
Knightime_X
Member
A: You assume, ai and what you can do with it has reached its limits / limitations. Lmao, Nope.
B: Missed the point, entirely.
C: Damn, if only it was possible to feed new data into something, developing if further....
ScHlAuChi
Member
Please define "AI" - I was specifically mentioning the current LLM based AI.
It is certainly possible that in the future they come up with something else!
No, I did not - an AI trained on videos is impossible to gain knowledge of the underlying programmed systems.
Please educate me where the data is supposed to be coming from.
Model collapse - Wikipedia
Even the french guy who invented this stuff says LLMs are a dead end:
AI pioneer LeCun to next-gen AI builders: ‘Don’t focus on LLMs’
Instead, Meta's Yann LeCun says, work on next-gen AI systems that lift the limitations of LLMs. But what are those next-gen systems?
venturebeat.com
IntentionalPun
Ask me about my wife's perfect butthole
Claiming AAA games next year is wild, but that’s Musk for you. Even claiming that this can generate any “arcade game” is pretty funny lol
It’s kinda neat but these look like games you’d have on your graphing calculator at school.
It’s kinda neat but these look like games you’d have on your graphing calculator at school.
Last edited:
Knightime_X
Member
For starters, never judge ai based on where it's at, but where it's going.Please define "AI" - I was specifically mentioning the current LLM based AI.
It is certainly possible that in the future they come up with something else!
No, I did not - an AI trained on videos is impossible to gain knowledge of the underlying programmed systems.
Please educate me where the data is supposed to be coming from.
Model collapse - Wikipedia
en.wikipedia.org
Even the french guy who invented this stuff says LLMs are a dead end:
AI pioneer LeCun to next-gen AI builders: ‘Don’t focus on LLMs’
Instead, Meta's Yann LeCun says, work on next-gen AI systems that lift the limitations of LLMs. But what are those next-gen systems?
5 years from now, this could be wildly different.
I never said that's exactly how that's going to work.
What I'm showing is, what you currently see will most likely look vastly superior in just a few years time.
The technique and method could easily change/ evolve over time.
You should inquire the people behind the games shown in the videos to get a better idea for the road map of things to come.
ScHlAuChi
Member
Where its going is heavily dependent on the underlying tech.
It will look sharper and have a higher res, but none of that changes the faults and limitations of the underlying tech.
Yeah It could - and in 2 years we could have self driving cars - but we heard that for the last 15 years!
And in 20 years we could have unlimited energy generation thanks to fusion - we heard that for the last 60 years!
Could, could, could!
I´m a game dev for over 22 years, and I tinkered with AI, as do many ppl in the industry.
We understand very well what AI is good for and what it ISNT good for.
And it certainly ISNT good for whatever Elon Musk or Phil Spencer promises!
Last edited:
BennyBlanco
aka IMurRIVAL69
ScHlAuChi
Member
Hey, you can believe whatever you want - time will tell who is right
Gorgon
Member
Knightime_X
Member
About cars..
Some people are on the struggle bus to follow 1st grade instructions at a self checkout stand, you really want them buzzing over your house?
You're viewing it in the NOW and treating like it's static and will never evolve.
The "now" for me isn't what I'm interested in, with ai it's all about tomorrow.
ScHlAuChi
Member
This topic was about Musk promising AAA games with his LLM based Grok next year.
Speculating about some far future tech that is 15-20 years in the future is pointless!
bender
What time is it?
To be fair, in Musk Time, 1 year is actually 10 years-never happening.
Unknown Soldier
Member
dorkimoe
Member
Elon Musk’s AI said he and Donald Trump deserve death penalty
Grok won’t answer the question anymore.
www.theverge.com
Grok does good work
ResurrectedContrarian
Suffers with mild autism
This is a simplification (the notion that synthetic data is insufficient, or even drawing a bright line between natural and synthetic).
There is still a ton of opportunity and ongoing work in creating infinite novel datasets today that propel models further every few months -- using synthetic methods, but decidedly not simply "feeding AI slop back into AI" or any other similar concept.
Instead, synthetic data is created by many complex configurations that leverage what we have available. A couple examples:
- diffusion-based models are trained on images distorted by white noise, where you simply add noise to existing images (or video sequences) and train the LLM to go backwards, to recover signal from noise; think about that pattern, and you'll see that it is used everywhere to give AI datasets of things that we can do in one direction, but we can't reverse. But the model can learn to reverse them. Give it video footage of the world and then take all the color out of it, and train it to go backwards (restore full color from black and white video), and it will learn it perfectly, despite that being a "synthetic" dataset. This extends to much more complex data perturbations happening across numerous fields, where you take data of some kind and destroy or distort it in way that AI can learn to go backwards.
- there are even more hybrid cases where we can go one direction but not the reverse, where I say "hybrid" because we actually leverage existing AI for both of the forward paths. Novel image editing models have been created this way. For instance, it's a very hard thing to find any large, coherent dataset of images where just one element is perfectly changed without any side effects (eg. removing a light source from an image, or removing a single discrete object but having all reflections, shadows, and other effects perfectly compensate). We can actually create that kind of dataset synthetically, though, by generated paired images through today's best image models, but manipulating them partway through, so that they diverge in a controlled way when creating the scene while staying aligned on everything but the one object we force to change. Then, you combine that with a series of modified (eg LoRA) adapters that give realistic varieties of natural styles on each image, to abstract over the synthetic image style effect. And you get a dataset that can train a model to take normal images (from your camera) and remove a single object or add a light source or other change, while photorealstically adjusting everything to fit, even the pose or lay of fabric and reflections to every pixel of the image.
Synthetic data is being used to great effect today, and leads to new capabilities.
Also, distribution datasets are not all that LLMs are trained on today! Reinforcement learning is making huge strides, and moves models beyond their training set. If you think "LLM-driven AI is just the dataset" this is false; it is now the case that LLMs can continue self-improving based purely on objective rewards (eg. did it finally solve the math equation, proof, puzzle, etc) and can churn until it solves, while updating its parameters. This is what was behind DeepSeek, for instance, and everyone is doing it now.
Even the french guy who invented this stuff says LLMs are a dead end:
AI pioneer LeCun to next-gen AI builders: ‘Don’t focus on LLMs’
Instead, Meta's Yann LeCun says, work on next-gen AI systems that lift the limitations of LLMs. But what are those next-gen systems?
Arguing from LeCun on the authority of his history is disingenuous, but unfortunately plagues AI skeptics.
Here's the simplest refutation (without having to go into why LeCun is a known idiot... he has said things about LLMs in past years that are simply hilarious on reflection in 2025, possibly the most wrong of all public commentators on the simple merits of the tech itself): Geoffrey Hinton and Ilya Sutskever (among many others) are notable for holding precisely the opposite view of LLM capability. And either of them has much more credibility as a founder of this tech. LeCun's contributions to today's AI stack are lesser in comparison. So there is no argument from authority; the best minds hold opposite views to LeCun.
Lastly, you keep thinking of LLMs strictly as language token models. That has already changed... their pretraining today is going from pure language to multimodal tokens from the ground up, seeing encoded image and video and sound patches, all modalities, and actually learning natively to relate the visual and language worlds. This--combined with reinforcement learning and model thinking patterns--means we are headed somewhere entirely new in capability.
lordrand11
Member
Ewww, get that shit off of Neofeed. Piss on AI for gaming, we need actual creative minds for it. Maybe throwing your money at new and upcoming talent will work better than eating a *consumed all this media* sized dog turd trope.
Pantz
Member
I've been using it for character battles.
Tifa beat Aya Brea. Tough fight, they both have high stats but Tifa's limit break was more powerful of a finisher.
Aya Brea actually beat Master Chief. He had trouble landing hits because of her high speed stats and her energy shot could take out his shield. Her liberate form was just strong enough to slash through his chestplate.
Aya Brea beat Kay Vess (Outlaws) with ease. The AI actually gave Aya a handicap by keeping her unarmed. Kay was able to land some hits and keep her at bay for while but Aya's stats are just too high, she closed in and knocked her out with her bare hands.
it's pretty entertaining
Tifa beat Aya Brea. Tough fight, they both have high stats but Tifa's limit break was more powerful of a finisher.
Aya Brea actually beat Master Chief. He had trouble landing hits because of her high speed stats and her energy shot could take out his shield. Her liberate form was just strong enough to slash through his chestplate.
Aya Brea beat Kay Vess (Outlaws) with ease. The AI actually gave Aya a handicap by keeping her unarmed. Kay was able to land some hits and keep her at bay for while but Aya's stats are just too high, she closed in and knocked her out with her bare hands.
it's pretty entertaining
cyberheater
PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 PS4 Xbone PS4 PS4
Enjoy these next couple of years before A.I really starts to bite into us. It’s going to be a very tough time for humans as we transition into the post AGI world. I’m not looking forward to it.
TheThreadsThatBindUs
Member
Humans... duh.
Errr, yeah. The code AI write will be in C++, same as what the human programmer uses.
Can you understand the text ChatGPT generates? Of course you can. That's the whole point of an LLM.
Errr... yeah. Neural Networks can use transformers and other techniques to train a model to understand the inherent relationship between the code it's writing and as many different aspects of the final output that you want to LLM to learn.
Wut?!?... That's what LLMs do by default.
And in game programming, much of the code that causes something to happen in a game is driven by functional, is standardized and entirely reusable. E.g. enemy pathfinding algorithms haven't changed for nearly 20 years. And AI would simply copy-paste the block of code for that.
ScHlAuChi
Member
Everything you wrote is absolutely correct, but none of this stuff will be any help to fully generate games!
But hey, I cant wait to see those big AAA AI games soon
Guess we will find out who was right in the future!
And just to be clear, I´m not an AI sceptic, I use it myself, but I dont just gobble up some stupid marketing hype by a serial liar.
DownLikeBCPowder
Member
Some are going to be pretty blindsided when AI is able to churn out decent games sooner than later, that's my conclusion from this thread. So many haters can't see what's happening because of their perspective, feelings, biases, or all of the above. I don't know exactly what it's going to look like, but AI is going to be absolutely transformative for gaming and the gaming experience. I'm both actively concerned about and interested in the results.
ScHlAuChi
Member
Ok, fine, but what do you think takes longer?
Reading, understanding and debugging auto generated code without comments, or rewriting it?
Just because one knows C++ does not mean that one can understand the structure of the code!
Sounds amazing! Cant wait to see the results:
AI is eroding code quality states new in-depth report • DEVCLASS
New research from GitClear, based on analysis of 211 million lines of code, suggest that AI coding assistants […]
devclass.com
Last edited:
Randall365
Member
Sounds like a platformer.
hemo memo
You can't die before your death
To be honest, if the choice is between a "modern audience" game and an AI-made game, I would choose the latter.
Last edited:
TheThreadsThatBindUs
Member
What makes you think the AI can't add comments? It surely can.
And yeah, for sure. Reading and debugging is infinitely faster. I dunno why you think otherwise. The point is, devs don't have to pour over every line of code the AI writes. They just need to check the parts that don't work or produce memory leaks. It's much faster than writing everything from scratch.
Errrr, knowing C++ means understanding the syntax, by definition.
That's precisely what NNs are great at.
I think you need to educate yourself on the capabilities of Neural Network based AI. I think you'll be surprised by what it can do.
Sounds amazing! Cant wait to see the results:
AI is eroding code quality states new in-depth report • DEVCLASS
New research from GitClear, based on analysis of 211 million lines of code, suggest that AI coding assistants […]
Lol, whatever that article is, it's either FUD or complaining about early AI models used to write code. The rate of AI advancement is almost incredulous. Give it a year and the state of the art models will be able to build full games in an engine like UE5 in a day.
