Thread 106164884

>Simply training the model to generate the next frame auto-regressively teaches it to maintain physical consistency across time
What's /g/'s take on this?

>https://x.com/AndrewCurran_/status/1952746390654009794
>https://x.com/agrimgupta92/status/1952735045527208016

can i goon with this?

>one minute
good enough for a road trip game

>>106164884 (OP)
I think this is kinda of a "solar roadways" dilemma.
Everything you do to make this better also makes other saner solutions better, such as just generate a gaussian splatting map in realtime and using the GPU to draw it.

>>106164884 (OP)
Yet another example of the bitter lesson that some retard holdouts still refuse to accept.
On the one hand it's absolutely insane that we can do this already. On the other hand, it will only serve to create more bland slop.

>>106164884 (OP)
>Generate a game where I have an interact function with my hand. This game is set in a middleschool of only girls, I am in the gym right now of that school.

>>106164924
this

>>106164884 (OP)
All aspects of human intelligence are emergent and will continue to appear in AIs as they scale.

>>106164884 (OP)
If you throw enough brute-force computing power at a problem, the model can even "learn" to encode the data it needs for some degree of persistence inside the frames it generates. WAAAAOW. So what do I get out of this? More AI slop?

>>106165010
Literally nothing, because moore's law is dead and you can't just "cloud computing" the way out this particular one.

>>106164997
>All aspects of human intelligence are emergent and will continue to appear in AIs as they scale.
Nothing of the sort is happening here.

>>106165018
we could literally be living during a hostile skynet takeover and retards like you would still be in denial. This post literally proves that we can still keep throwing more compute at the problem and get actual results.

The fact that this stuff doesn't even trend anymore tells me that something killed the hype

>>106165054
You can, but i will toss compute in a smarter way, and make a skynet that will eat your skynet for lunch.

>>106165063
imagine caring about muh ""hype"" in the slightest.

>>106165054
>. This post literally proves that we can still keep throwing more compute at the problem and get actual results.
"Actual results" being what? More ugly slop? If only we could heckin' harness the power of a black hole and convert the entire solar system into corporate computing centers, I bet trying to brute-force intelligence will finally yield useful results.

still playing genshin impact

>>106165063
>The fact that this stuff doesn't even trend anymore tells me that something killed the hype
Because WHERE ARE THE PROFITS FUCKLEY
AGI IN 2 MORE WEEKS GUYS
THROW ME EVEN MORE CASH YOU STUPID PAYPIGS
ISN'T IT "INTERESTING" HOW EVERYTHING WITH AI "FEATURES" KEEPS GETTING WORSE

>>106165093
I hate ai slop as much as the next guy but you have to be completely braindead if you don't see how this is impressive.

I was worried about the Jews having my penis pics but they will just use this to make some incriminating shit against me with gchq signing off on it.

>>106165063
I wish, then all the fucking shills could disappear and leave the people who are actually interested in the tech.

nintendo is finished

>>106164884 (OP)
it was trained for consistency. i guess they recorded gameplay clips with input commands and feeded it to AI. there is consistency in gameplay clips.

>>106165063
Everybody is saturated to death of AI slop thanks to YouTube.

>>106165123
It's only as "impressive" as you subjectively want it to be. What meaningful new capabilities have been gained? What new understanding has been gained?

>>106165123
It is pretty impressive, but people thinking that this specific tech can be used to replace video game making are A) out of their mind, and B) not seeing the 20 actual ways AI can actually do it.
You can just create an AI friendly game engine and do one right now that pulls it off.

>>106165202
so the idea is that inconsistent predictions are rejected. so it's not just next frame prediction

>>106165268
>it's not just next frame prediction
This is the same magical thinking as "it's not just next token prediction" where by definition that's literally all it can possibly be.

>>106164884 (OP)
use case?

>>106164884 (OP)
Oh wow, it fucking remembers trees? I'm gonna drink the kool aid nao!!!

>>106165283
there is a hard time limit. why? if it was just "next frame prediction" this limitation would not exist.
because it's only has a temporal consistency. it tries to predict the whole recorded gameplay clip.
so this approach is not scalable.

>>106164884 (OP)
Wow, a whole minute? And how much energy is this thing hoovering down to remember where a tree is slightly longer than a goldfish would?

A 3d renderer does the same thing, why are they trying to use neural networks for tasks where better solutions already exist ?

>>106164884 (OP)
>across time
For how much time though... autoregressive = quadratic time = it will crap the bed eventually

>>106165450
>>106165458
It's not about use cases. It's about sending a message. Don't you feel the emergent AGI magic?

>>106165458
Can you prompt a 3d renderer and get a whole simulation without any code?

>>106165598
No. What's the use case for "prompting" some generic scene you have little to no control over?

>>106165629
It looks cool?
Obviously they aren't going to stop here and will put more effort into evolving the technology.

>>106165657
"It looks cool" is not a use case. Even with static image generators, it's basically impossible to control the result beyond broad strokes. People who use diffusion models for professional work end up convoluted multi-stage workflows that usually involve stitching together parts from different gens and manually editing everything into coherence. Doing that for a real-time clip is impossible.

https://youtu.be/qC5KtatMcUw?si=B5Unna_l9NHS7lPl&t=157

Genieception

>>106165629
training robots in diverse environments without having to code them all up manually

They didn't really advertise this in the announcement, but the real purpose of this model is to train real world physical robot agents. It can take any video as an input and then extrapolate it in real time while also adding new inputs.

>>106164884 (OP)
Can it make coom or should I ignore it like most other AI slop?

>>106165121
>AGI IN 2 MORE WEEKS GUYS
*years
https://ai-2027.com/

>shill link
added to the filters

>>106166156
>training robots in diverse environments without having to code them all up manually
It's much cheaper and more controllable to use procedural generation.

>>106166185
i really hope that's not true, because semi-convincing object permanence in video generation is not gonna give you the ability to inspect distances and clearances in the map (radar / lidar ranging, cliff detection, even pose), and without trustworthy validation, any "training" of the perception stack is suspect at best. shiggy dig.

>>106166302
>because semi-convincing object permanence in video generation is not gonna give you the ability to inspect distances and clearances in the map (radar / lidar ranging, cliff detection, even pose),
Humans can do all that based on video. We're not asking the machines to perform magic here, just something that we know is already feasible.

>>106164884 (OP)
It’s more overhyped LLM tech. The same thing just repackaged - the impressive thing is the speed at which it can generate frames, but that’s presumably just because it’s got vast compute powering it (it’s a google research project, they literally have the worlds most powerful super computers). Imagine this compute powering a raytraced ultra max settings traditional video game.


The object permanence seems amazing at first, but it’s basically just a context window in a visual sense- it’s like clicking around a google streetview but it generates a new frame 24 times a second. Of course moving forward will make things get bigger, it’s not actually that magical when you see what they’ve done, it’s impressive, and a clever trick but you can see how it works. It basically retains every frame it generates in memory and can retrace its steps if needed. But it has all the same flaws every other LLM has, it’s baked into a single canvas/can’t tweak it/massively inefficient/hallucinations everywhere/huge compute cost/generic slop outputs that never beat human engineering/artistry.

There was a version of Doom they had running on Genie, they trained it on 30,000 hours of Doom footage. They spent a trillion dollars getting Doom to run on a super computer and it looks like shit.

>>106166330
Humans can also count the number if 'r's in strawberry

>>106166360
Some humans.

>>106166330
prove it. give me a plot of altitude with respect to time from that video, and a list of eigenvectors to feed into a 9-dof.

>>106166338
Also forgot, I HOPE YOU LIKE MASSIVE DEADZONES AND LAG IN YOUR CONTROL INPUTS! and yeah if people thought the artifacting and smeariness was bad with modern AA then just you wait for this shit fucking warping all over the place

>>106166399
I would love to, but my bigger than average brain is a black box, sorry. I could demonstrate this ability indirectly by operating a drone from a video feed and not crashing.

>>106166446
no worries, i accept your concession that the source material is not adequately usable.

>>106166457
Enjoy your hallucinated victory.

>>106166465
be sure to drink the entire cup!

>>106164884 (OP)
please please please please please jewgle give me the blue bill fit me in the matrix pod I don't want to live in this 3dpd world anymore milk me dry do whatever you want but make me live the good life in the digital world

>>106166481
>implying the corporate digital world would be good
>or even fun
Did the absolute slop that the metaverse was and the nonstop talking and whining about "AI safety" not clue you in? You'd be unironically better of just going on Second Life

>>106165063
it's not hyped because it's currently just a whitepaper, less than a tech demo
there is no product yet

>>106165123
A little but what is the actual use case?
>playing game
>get a text, answer it
>turn around in game and everything is different
Maybe a race the sun type game where you only go forwards but that can already be done without using absurd amounts of power. The problem with AI is the data center needed to run it. AI-bros seem to ignore the real world costs to make this slop work and only focus on what could be if only we wait FIVE MORE YEARS.

>>106165284
yeah this
it's yet another WOW SO COOL then you try to think of a use for it and nothing comes to mind
they've been doing this for years now, making impressive looking stuff with absolutely no application, and morons are so wowed that they throw money at it

>>106165063
wipe your own ass

>>106164924
the only question that matters kek
it's google though

>>106165063
it's a cool concept but then you also realize that it's closed source saas shit that will prioritize "safety" over all else

>>106166481
This but unironically

>>106167128
>then you try to think of a use for it and nothing comes to mind
It's like anything procedurally generated but this time, no code. Program someone's dream, code a sequence for a TV show, maybe a little game-in-game when the main game loads.

>>106164940
Honestly I thought this was basically what it was doing. Doesn’t really feel like new tech to me after playing with splats for awhile. Just faster and more spread out. Feels like they are sidestepping the knowledge / memory side of it because they have no clue on that front.

I wonder what's stopping them frmo using a different kind of input like say, VR hand controls. The idea alone of being able to interact with a purely simulated world rather than just looking and walking around is huge.
Thinking about it from a game perspective, it would mean no more traditional game physics but instead realistic consequences to your actions. For example, pushing someone wouldn't trigger a ragdoll effect or a stagger, but instead the person being pushed would behave like an actual person being pushed around. On another note, I wonder if the "violent video games are bad" would make a return in that case lol, considering the simulation would be almost as real as the real thing.

>>106167622
Anon, I think it already can take almost anything as an input because its basically text2video that run in real time and can be edited in real time. But I doubt it will replace traditional games because of how expensive it is to run, and because games are more about gameplay and set rules then just pretty graphics and simulating reality. I doubt gamers would tolerate AI hallucinating if you got shot by an enemy or not, all the previous iterations on this like the DOOM, Quake and Minecraft allowed you to walk around fairly well, but then shat themselves when moving entity popped in.

There's no way we can have this run locally with even expensive hardware of today right?

Long ago there was a market with MMORPG where this kind of technology to do open worlds that change would make sense. Now? What the fuck is the point? The MMORPG genre is defunct and the ability to have AI handle environmental regrowth, and changes if someone diverted a river has no end utility. 20 years ago people would have shit themselves at having AI able to dynamically simulate random orc tribes moving around or manifesting out of nowhere on some frontier region, now its a "meh" because there is no way to exploit these AI tech for anything beyond a tech demo.

>>106164924
Not with a circumcized dick

>>106165018
If you get cheap mass produced reliable AI intelligence, progress and experiments can run much faster. With enough AI personal, you'll see much more research. It'll cure aging, injuries, everything.

>>106164884 (OP)
Schrodinger Tree is SOLVED.
/g/entlemen, like the tree, let us remember where we were, for a minute, on the day this remarkable breakthrough for humanity was pronounced.

>>106168053
> It'll cure aging, injuries, everything.
Lol what? How so? This is baby level thinking.
>if I have baba all problems are better
You are actually retarded and need to seriously ponder the way you process reality.

>>106165063
Because it's pointless garbage. Anyone with a brain is already tired of slop with zero intent or purpose to it.

>>106168336
i want it to invent women that talk to me instead of running

>>106164884 (OP)
>visual memory extends one minute
Oh, so still fuckin useless then, neat.

>>106165458
The benefit will be consistent video edits with less tripping
3d game usecase is shit though, agree, unless it fully relies on video streaming as a game, but that would be too janky, too expensive for both sides, and too inconsistent that even Kojima wouldn't make it good

>>106171737
The developers themselves call it a world model not text2game or game model. Their goal is to use this to train robots on 1 minute randomly generated tasks and maybe help them more with spatial awareness. Only the biggest retards on the internet look at this and think this is the future of videogames.

>>106164884 (OP)
>maintain physical consistency across time
>for ONE MINUTE
usecase?

>>106172300
Tech demo shit and for training robots, or as they call them "agents". Its also designed to be a world model that AI researchers talked about half a year ago where instead of making AI just predict the next word and then make it learn from there, we make it learn how objects in real world behave and give it spatial intuition instead of just hoping LLMs will develop it on their own. Its basically for them one more step towards AI that can actually reason, but aside from that its just a fun gadget that is too expensive to release to public.

I'm a robotics PhD student and this makes my dick go hard for the future of me stealing your job

>>106172459
You wont steal anybody's job. You will just work on one of many robotics companies making robots which will then steal jobs. With robots stealing manual jobs and AI office jobs maybe there will be enough unrest to make or dividends redistribution happen.

>>106165063
To the average person it isn’t all that different from video generation, it will get hyped once it’s capable of making a demo of a Skyrim like game

>>106172459
Do you not worry that your creations will in turn steal your job?

>>106166185
imagine all the ziggers that will die by autonomous drones trained with this shit

>>106172842
Robotics is for manual labour. As a phd graduate working in some robotics lab he won’t have to worry about that. AI researchers on the other hand are the ones making their own replacements. Recently there was this Chinese model that was able to make better more optimised and more reliable small model architectures for like one class of small models. It is a pipeline where it goes through the scientific method to achieve the goal.
>>106172825
Idk what you mean by Skyrim demo. Is it just the player walking through a grey forest and talking to random NPCs on the way or do you mean actual gameplay that isn’t clunky mess? From the fact that video models and even image models suck really hard at action scenes and combat I am guessing it will take some time. Programmers who aren’t coding some high end operating systems or brand new graphics rendering engines will probably be gone faster then this shit learns how to do combat.

>>106164940
And how exactly does your "map" (subtle pedo reference btw) give you items that can interact with the world such a the paint brush or bicycle in the demo?

>>106173204
> Idk what you mean by Skyrim demo. Is it just the player walking through a grey forest and talking to random NPCs on the way or do you mean actual gameplay that isn’t clunky mess?
This webm was done with the help of AI, though required more human input than a simple prompt which is what Genie is going for, and managed to normies pretty hyped

>>106172459
the more LLMs displace knowledge workers, the cheaper the costs of manual labor
and that in turn makes robotic automation less competitive

so they wasted enormous amounts of energy and space just to force their garbage to remember 1 minute window. truly amazing tech. can't imagine why it's not profitable.

>>106166067
It's fucking over

>>106172545
Yeah that's what everyone in the field says so that must be wrong
>>106174971
Interesting take

>>106172459
I'm the doctorate level Robotics LLM that's hard for the future of me stealing YOUR job

>>106174895
my dream as a kid was to walk towards the hills, mountains, terrain, castles in the background skybox
AI will make my inner kids dream come true

>>106174895
>no anti-soul gas
unplayable by modern audience

>>106179533
Have you played literally any open world game in the last 15 years?
>>106174895
This is something you can make already if you just let AI generate you all the sprites and you decide to make your own better looking daggerfall.

>>106168021
im fine then

>>106179772
>fog
>volumetric lighting
>invisible barriers
>ice trolls

>>106166067

>walks out the window as if it's a door
>turn around
>the window is now as wide as a garage door and the room dimensions change

there's no consistency problem solved

wtfs with the mental retards shilling ai all of a sudden?

>>106179772
>Have you played literally any open world game in the last 15 years?
yes, and it's disappointing
every game is full of visible and invisible walls
>you see that building over there?
>you can't enter it
>you see that island over there in the distance?
>that's only a texture
>you see that ominous castle 100km away?
>don't even dare to look at it longer
I hate how restrictive games are
why do I have to move left or right? I want to explore the jungle
I want to move towards the trees