Quake 2 was developed by 13 people on a 1990's computers.
The paper cited by Microsoft lists 22 authors and, from what I understand, claims that results are consistent after a Sextillion of FLOPS. (that's over 3 megawatts of power if ran on a 5090 if I got this right)
https://www.nature.com/articles/s41586-025-08600-3#MOESM3
The paper cited by Microsoft lists 22 authors and, from what I understand, claims that results are consistent after a Sextillion of FLOPS. (that's over 3 megawatts of power if ran on a 5090 if I got this right)
https://www.nature.com/articles/s41586-025-08600-3#MOESM3
Comments
Using unimaginable amounts of resources to make something that can be made without AI, but worse. What does this prove? Is this the flex AI bros think i is?
I am so very tired of this BS.
But would they publish this in this sad state if it could be easily improved? ChatGPT still can't do some stuff it couldn't do when it was released. Pictures still have artifacts. Really think it can still get better?
Here we have an AI prompt create an entire game - that is IMPRESSIVE. Like SpaceX's landing rockets. Is it flawed? Of course!
Do all of you hate Star Trek's Holodeck? Because this is that in its earliest stages.
Maybe you'll turn out correct, but I don't understand the negativity. This is exciting.
What you CAN do is, you can have it summarise and rephrase longer parts into shorter, introducing more errors.
There is no magic. It's just a prediction machine. "What's the most probable next frame?"
What do you think happens when you play any game? ;D
Look, nowhere am I saying this is currently perfect. But it's the start and even if it ultimately fails, I'm willing to give the potential Holodeck-precursor a chance.
I mean, we're generating text using floating point data out there.
This isn't "as far as it can go" this is already much further, just brute forced with $$$
We have so much room to go on.
1. Garbage
2. ???
3. Not garbage
This tech is going nowhere, there are things it will never be able to do and there is no way to fix it.
It’s basically a „million monkeys with typewriters” situation. It does not generate text and it does not draw. All it does it predict, all it generates are probabilities.
Can we please instead start focusing on its current state for once, combined with clear-eyed technical analysis?
Nearly all tech has some bigger or smaller thing that can still be incrementally improved (at least for a specific use case).
What matters primarily is whether the fundamental problems are at all solvable. With lossy solid compression they're not.
They can't make this without making the actual game first.
They can't (statistically) compress untold amounts of data into a much smaller fixed amount of space without losses.
They can't escape combinatorial explosions.
The AI is worse *by design*. And that pisses me off 🤬
I had it done somehow on my cellphone, but it's swiping from the corner of the device instead of holding the back button
( I do agree on your general point tho)
it's not just not functional and messy, it's actively wasting resources on a worse version of the future
It's 10 sextillion FLOPS not 1 sextillion.
Using appropriate hardware (not gaming GPUs) it was probably $350-$500 of power draw.
The paper:
"Models towards gameplay ideation"
The only reason I haven't blocked you yet is because I'm looking for a half decent labeling system.
https://bsky.app/profile/northernnerd.bsky.social/post/3lma4bwdtrc2x
"gaming is the entertainment industry’s largest sector worldwide"
Or in techbro speak: 🤓💲🌍🔥👹
There are really groundbreaking uses for this technology, but genAI is not one of them in any form.
Am I going to play this as a simulacrum of Quake? No.
Did I think it was neat regardless? Yes, absolutely.
Moreover I don't think your claim it took exponentially more resources is even accurate or defensible.
Marketers being disengenuous, which is a valid complaint whether it's AI or literally anything.
My point stands they were never intending to recreate quake.
This is rubbish.
An RTX 5090 is said to provide ~419 FP16 TFLOPS and has 575W TDP.
You divide the sextilion by RTX flops to get how many seconbds are needed, then divide that by 3600 and multiply by the TDP.
Still, I might be off, and the industrial-grade AI chips might have better FLOPS/W rations but It's still an absolute shitload of electricity.