i built the 1st "world model game engine" it makes games from scratch, all running locally on my Mac M1 - 30fps - 40ms latency - ⁠50 million parameters - trained on 15 min of footage made using only "synthetic" datasets this is the first of its kind, &... you can try it nowshow more

this is not a game engine

Wow, I tried your "game engine" and it's just a bunch of pre-recorded video sequences that play very slowly when I press keys. You pretend you reproduced a bunch of games but you didn't even do that. You just reproduced some frame sequences. Btw this is its real speed & quality:

hey! the model run on-device, so performance depends on your hardware. What GPU do you have? But it's generating frames in real-time, not playing video, so if your framerate drops, it feels less reactive. That's the tradeoff when you're running inference locally :')

I'm not sharing my specs but they are irrelevant. You're showing seconds of pre-recorded video at a 1990's quality. If you need special hardware for that, better go back to the drawing board. And if you want to make a "game engine" have a look at e.g.

Needing an M1 or higher end GPU to run this pile of shit on a browser at 30 fps with 40ms latency is completely missing the point of what a video game is. I have nothing against AI, some even unfollow me for being an “AI bro”, but this does nothing for video game development. It’s interactive brainrot, at best.

@hugothomel Yeah, AI folks constantly try to redefine terminology to suit their needs. This is first of all bad for AI research. It makes it all look more like a scam, and researchers like used car salesmen, than anything genuine. Btw, I'm an AI researcher. Like I publish in AAAI, MLJ, etc.

@hugothomel Which, you know, makes it sort of my job to be critical 😅

join the creators waitlist → play the demos →

crazy, i have a couple game ideas. When can i try the game maker ?

join here for early access! we're opening the platform in a few weeks!

Hey cool! We built the fist world model game engine at overworld! Should find a way to collaborate :)

this is so sick! found the link:

haha you even found the ugly link! you can also use that

goat @JustJake out here powering out world models

Show me again a year from now after you've put serious work into it.

stay tuned! I can tell you that what we'll have in a year will knock today's demo out of the park

Hey bro, the community loves your work so they created a fan BAGS coin where all fees are directed to you, check it out! 7vm5eNQTvFWrNkWaEjV5L8nrmxPLGWbaMqA3PPXpBAGS

If you need angel checks or early users I'm happy to be both/either. DM me!

thanks Astrid! I'll dm you I'd love to have a chat about both!

Very cool! Reminds me of our first model we trained on 5 hours of me playing a single game. Out of curiosity are you finding that the model is remembering things as they go out of context length? That was one thing we saw back then as it started memorizing 3d layouts implicitly.

Yeah, actually! Increasing context length yields improvements but it seems to plateau quite fast. And there's fine line to walk to not fall into overfitting, especially with a limited dataset. Would love to chat more about that 👀

Yeah for sure! Happy to move to DM. From a research angle I think there's a lot of unique capabilities your approach has that traditional large WMs might not.

Let's do that! Very curious to have your take

Pretty amazing.

I convinced we're gonna see an exponential rate of improvements from here. You can see this as the 2023 "will smith eating spaghetti" moment of world model gaming!

Agree.

got an error when trying doom

Oh the doom one is insanely more demanding to run! I should just go ahead and remove this actually, the clones there are very old projects haha

lmao how about you go fck yourself with your cute stupid video generator

I think I have some pieces to make this even better. Let’s talk

hey I'd love to! Shoot me a dm and let's setup a chat

bro really made this tech just to post "first"

Me thinking this was the main driver for all human engineering progress😅

yo pas mal

merci ! and we're only getting started 😎

what's your timeline for UE5 replacement it's immensely important

Hey @hugothomel the community saw this and made a token for u on @BagsApp to help u with funding!! I sent u a dm with more detailed info!!

Sounds like a pile of shit, will keep watching though

the background music 🔥

big shoutout to @super8n for tackling the challenge of helping me with this video 🔥 ML has nothing on communication haha

Sorry but you haven't made anything that works, but look forward to the next attempt

very cool! a core criticism for world models in making games is that they are far too non-deterministic. need rules for mechanics, systems, goals, etc. even tho you can use synthetic data for new mechanics, how are you thinking about that?

That's actually the core advantage of our approach. Non-determinism comes from inconsistent training data. With synthetic data, you control the distribution, every example shows the same mechanic, the model learns it as a rule!

@numberdust say I’m playing a good run and I fall through the floor and die. can you address that? that happening a few times can make someone put a game down forever and if it’s an engine thing thus make a dev put an engine down forever

@numberdust that's a very good point! so what to keep in mind is that the dataset is your code. if you don't have examples of your character falling through the floor, it's not likely to happen. But let's say it did, the fix would be to add counter-examples in situations where this happens

@johncoogan @jordihays @RealProducerBen

lowkenuinely scary

not as scary as the games that’ll be made 👀

Oh wow, this was very informational. I look forward to the future with the highest hopes I've had in a long time. Thanks to genius brains like you haha.

We see eye to eye! The purpose of this engine is to augment, allow creators to make things that were not possible before. What I always loved about gaming is how it bridges art and technology, so providing a new medium of creation is total peak for me

Imagine the possibilities, not just for games/art but the education aspect of this would go crazy to be honest! So many applications and so little time. I'm working on a little something of my own atm. Hopefully here in the coming months I have something to show for it!

Absolutely! We are razor focused on games right now but I think it can really enable a shift in the entertainment industry in general. And edge computing remains an unsung hero of the modern age haha

pretty sick

And I think the applications go beyond gaming. When you see experiments like Bandersnatch, I wouldn't be surprised to see the line between cinema and gaming blur as time passes, don't you think?

i believe this is how the universe works ;)

Nicely done! Looking forward to the @GetSoloTech collab 💯🦾😎

@GetSoloTech Let's get this done! I'm at if you ever come around, but I would love to chat

Great work !

@Param23072004 thanks Param! early days but very thrilled to share this! Hoping to see you in the beta 🙂

For sure this looks awesome and I’d love to try it. I’m really curious how you developed it, especially the system design and architecture, and how you’re getting this kind of output with just 50M parameters. Would love to chat to learn more about it!

running on an m1 at 30fps is impressive. curious what the training data looked like - game footage or something else?

It's actually purely synthetic data obtained from video models, that's how we can create games that didn't exist before!

50 million parameters? no way lmao

I know right? Took a lot of iteration to find the right balance, small enough to run locally, enough capacity to generalize. The trick is what you leave out!

you plan to reveal more about this? I kinda want to try this on very small games

asks permission to connect to ANY device. haha. nice try.

🔥🔥🔥

Let's go!

I love it! Looking forward to giving you access man! Any game ideas yet? 👀

Heck yeah! Would love to make something similar to Final Fantasy series or even similar to the PSP versions

Man I love the ambition! If I were to do it right now I would maybe train several little models instead of a bigger one. Happy to chat more about all that when we go live!

Definitely! I will be eagerly waiting for upcoming updates and waitlist access!

Awesome! And out of curiosity, have you been experimenting with similar tools as well?

Yup I have tested some other world models in the past, but this being world's first of it's kind, it will definitely be an amazing experience!

You'll be creating your own tiny world models!

The fact this runs entirely on-device is doing most of the talking.

when I realized there was a way to do this, it was the last piece in the puzzle that clicked in my head!

Pourrais-tu nous faire une petite vidéo en français BG parce que on a des grosses pépites de l'autre côté de l'Atlantique

haha avec plaisir ! j'ai toujours rêvé d'avoir une excuse pour utiliser @synclabs