Moonlake AI raises $28M seed to build interactive AI-generated worlds and games

to getbasel. com. Your bezel concier is available now to source you any watch on the planet. Seriously, any watch. We have our next guest from Moon Lake in the TBPN Ultra Dome coming in from the reream waiting room. Welcome to the show. How are you doing? Welcome to the stream. What's up? Thanks for not too much.

Uh I we we got to be fast about this because I want reactions to all the stuff that's happening in the timeline. But uh kick us off with an introduction on yourself and the company. Yeah, I'm Fannon. I tend to go by sun. I'm the founder and CEO of Moonlink AI. We're really spun out of Stanford AI lab.

Um, a couple around six months ago, we started this company really to build AI that can generate real-time interactive content, think simulations and games. Yep. So, yeah, I mean, what what are your reactions to Sora 2 vibes? There's all these different paths that are people are testing. Is it a tool?

Uh, then there's also this uh interesting take about maybe we won't even be consuming it's all reinforcement learning data for robotics. like how are you viewing the market play out just in AI generated content, video, virtual worlds, all that.

Yeah, I think it's very apparent that there's a lot of people going through this space and it's it's really just because the the potential is unbounded, right? Being able to model the world, let alone any other virtual reality is going to the biggest the next biggest wave in AI, no doubt.

Um, now the what we're doing is actually very different. Yeah, I was going to ask video. Yeah, video video. A lot of people playing an image and video. We don't we haven't gotten a lot of pitches around creating sort of these interactive media. Something more like you could imagine a mobile game built on Moon Lake.

Am I am I going in the right direction or totally? Yeah, you can imagine games being built simulations. And really the main difference when it comes to simulations and games compared to videos or images is the interactivity, right?

User outcomes change the trajectory of the world or the game or the environment that you're acting in and that's what really differentiates us from videos or image generation companies. How do you think about world? Yeah. How do you think about tool use with uh chat GPT just the knowledge retrieval app?

We went from GPT3. 5, it was okay. GPT4 was great. Then we added tool use and we and there was a moment when it was like just keep scaling it up and it'll learn every the answer to every math equation.

And then the engineers open AI just realized like wait let's just teach it Python and it'll just do the math in Python and then you'll get the right you'll get the right answer and you don't need to memorize everything into the weights of the actual underlying model.

uh and I when I see these uh virtual worlds, these generated worlds, I feel like we are in the GPT3 moment and we haven't we haven't yet given them a database to store your inventory or given them the ability to even a notepad to write down what happened in act one of the game.

How do you think tool use comes into these generative worlds uh in and will it track with what's happened in the chatpt app? I think tool use is exactly the right paradigm that we should go about thinking about creating virtual worlds.

You know, computer graphics has been around for decades if not like hundreds of years, right? Um, and we have the tools that enable us to create the highest fidelity virtual worlds already, right? Look at all the movies, look at GTA, right? It's just that these things are hugely expensive and manual to create today.

Like GTA costs $2 billion to de develop, right? Um so really what we believe is that we should use reasoning models to reason about the representation, the tools to use, the models to use and consolidate all these things into a platform that allows anybody to be able to vibe code their interactive worlds. Yeah. Right.

These are the things that are historically just barred by um just complicated tools and and domain expertise. But it doesn't have to be that way. Do do you think do you think uh how how do you feel the legacy game studios have been reacting to the opportunity in generative AI so far? Have they been? Yeah.

like have they been I I you know I don't follow gaming super closely but for example like the the EI EA take private you could imagine they're they're they start huge opportunity for sure they they have a lot of IP they could if they're private they could have the potential to take a longer term view on AI and try to reinvent themselves but I'm curious how you view it.

Yeah. No, they're all the studios that we talked to are stoked about it. Right. So, we were actually talking to a team that owns F1 and they were talking to us about the possibility of being able to create like racing games that ghost races with F1 cars in real time, right?

And I think what's what the way we should be thinking about this is that what we're building at Moonlake allows things like so many more things to be built or things that are historically cut or just made um impossible due to the time and resource that it entails to create these things, right?

And so really all the studios and um folks that we work with on the gaming side are stoked about what we're building and there a lot of them are you know looking to collaborate with us. What are your what's most rate limiting to you right now? Is there enough data out there? Is there enough compute out there?

Is it just do you need to you know just go and have really smart scientists think and think and think and come up with an entirely new uh you know technical paper an algorithmic innovation like what is on your critical path? There are definitely a couple of algorithmic breakthroughs that we are still working on.

We have some of the best scientists here to to work on those problems.

But really even without just like significant breakthrough I'd say the technology today if you just consolidate them with reasoning models the upside is already unbounded right look at Roblox it's a $100 a hundred billion dollar business with really more traditional graphics approach compared to like all the generative world model techniques that has I don't know if you're familiar with this Ben Thompson take, but he he was saying that like chatbt is so successful that at this point OpenAI could stop training foundation models and it could be clawed under the hood and it would still be a fantastically amazing business because Chacht has won the application layer in knowledge retrieval.

It's the app on everyone's phone, right? uh like do you see yourself becoming an application layer company staying in a foundation model layer company like doing some sort of blend like how do you see your business developing?

Yeah, we are actually an applicational company so we are going to build products that allows anybody to be able to create interactive experiences. Yeah. Right. The way we go about thinking about oh whether we're product or model company is really that product is first principle. Yep.

But a lot of the times given what we're building is at the frontier, we need to do a little bit more research and some model training in order to power the best product experience. Yep. So the analogy I like to make is is cursor, right?

Cursor didn't start off training their own foundation model, but they had their tab completion model. They had these smaller modules that come together to build to form an amazing product that people are people love, right? And that's what we're doing. How crazy is the inference cost right now?

It feels like Sam Alman's setting the GPUs on fire with Sora 2. Uh obviously per token inference cost is declining by a thousandx every couple years or something like that. But uh for all the demos that I've seen in world generation, uh they've all been somewhat limited. It hasn't gone super viral.

I haven't seen, oh yeah, there's 100 million DAUs on G3. It's like Genie 3 isn't even available. And so uh how how GPU constrained how crazy is it to actually go and uh release one of these things into the wild like and what's the path for you know actually bringing down the cost of inference here. Yeah.

So I would say inference is definitely something that we think a lot about and compute is definitely something that we use a lot for. So if you're a GPU vendor you know I like to talk. Okay.

But I would say our unique approach of using code to represent the world provides persistency and it consolidates all the tools such that we can use server u client side compute um to calculate next state of the world. Okay. Right.

And that actually allows us to not have to spend a insane amount of compute on just generating the next frame of the world. And that's because of our unique approach of combining co-gener co-generation models and diffusion models. Right.

The way we're building the way we're building this product is using a representation or a hybrid representation. Yeah. That allows models to choose what representation to use to represent this particular problem in this particular environment.

And that actually allows us to be very computationally efficient compared to existing role model companies or video generation companies. Yeah, there's been a ton of push back uh against AI video generation. Like the reaction to Metavibes on X at least was very negative.

Uh people were pretty amazed by Sora 2, the model, but then there's been a lot of negative takes about Sora the app. How do you think this will be received once we have a vibes or general availability just an app that anyone can download and start using really early? Do you think people will be more optimistic about it?

Do you think there'll still be this push back and then you have to get through it somehow? How are you thinking? My my view is the the viscerally negative reaction that I have to a slop AI video is because of like the character inconsistencies and like something about it just feels like very wrong to look at.

Whereas with Moon Lake, I imagine you can create this like gen, you know, if you generate a a consistent world that's interactive and it's beautiful. Something like being in a midjourney. Yeah. Prompt that I think would be it's about the experience the user has when they actually are playing around with it.

And like today, Sora can be entertaining, but still you can have a negative reaction to it. Yeah. Whereas I can see this done correctly. Yeah. H how do you think people will react to when this technology broadly probably from you hopefully uh becomes general availability to the point where every Yes, I like it.

Bet on yourself. Um the to the point where it's just an app, anyone can download it and like we're seeing millions of people on board. What do you think the reaction will be? Oh, I think people will love it, right? Because what we're doing is not like we're generating the content that people consume.

We're just putting great tools into people's hands and allowing them create and craft the the experience and worlds that they desire, right? And we but we're just making it 10x easier, 10x faster, 10x more accessible. Take us through the round and I also want to hear about the uses, but give us the numbers.

What' you raise? Also, is this your first company? It is. Nice. $28 million seed round. You're in the big leagues. Straight to the big leagues. Congratulations. Uh yeah. Who did the round? How much was it?

Yeah, so we raised 28 million from Ventures, AX Ventures, Nvidia Ventures, and alongside 10 plus unicorn founders and angels and some of the most influential AI researchers, too. We're super grateful for their support and I can't wait to see what people build on top of Moon Lake and play around with it.

And uh I'm sure we'll see you back on here soon. Thank you so much for joining and and congrats on that. Congratulations. Thank you for having me. Cheers. We'll talk to you soon. Up next, find your happy place.

Book the Wander with inspiring views, hotel, great amenities, dreamy beds, top tier cleaning, and 24/7 concier service. It's a vacation home, but better. Uh, you know, I I one day you're going to snap when I cut you off. I still think it's funny.

I think some people are getting sick of it with the with the ad transitions, but uh we're having fun out here if we have some news. I have some news in the chat. Uh they said the the the circle table is growing on them and the horse in the background is growing on them. They say never.

Can we get a can we get a shot of the horse, please? Our beautiful horse. This massing it. This this uh you really have I