Moonlake AI's Fan-Yun Sun is building interactive AI-generated worlds using code as the world model

let me tell you about public.com. Investing for those that take it seriously. They got stocks, options, bonds, crypto, treasuries, and more with great customer service. And without further ado, we will bring in Sun from Moon Lake, Sun and Moon. You see what I did there? Congratulations on the launch. How are you doing? Thank you so much for joining joining the show.

Um, and uh, please uh, explain exactly what you built and how much AI is going on because it's a crazy demo. We're hopefully going to be able to pull it up, but we'd love to know about how you're positioning the product, how you're explaining it, and what Moon Lake is. Yeah, of course. So, actually, let me take a step back, explain.

Define role models first because there such a term that is overloaded. Yeah. Um,

so I define role models as

models that can ex a predict a model that can extrapolate the next state of the world in an action,

right? And in fact, I'm defining this right now because I'm predicting that you guys are going to ask me about okay, how is our warm up different than other warm?

Sure. Sure. So this capability is actually just like my world model capability.

Yeah.

Right. And biggest the biggest problems with foundation models and AI today is that these models don't have these world model capabilities such that it can't predict and act in long horizon tasks.

Yeah. No, we've seen with like Genie 3 amazing ability to paint on the wall, come back, paint again, and the paint's still there. But we're so far away from a real game mechanic of you're 5 hours in and you go back and you get a new quest. Like we're we're far away from that. At least it felt like that until this demo.

No, exactly. Um like there's there's different schools of thoughts when it comes to role models

and really it comes down to how are you representing the world for the tasks that you care about simulating, right? So for example for Genie um it's they have really pretty pixels. Um the problem with with it is that it you know can't remember the world currently more than a minute.

Yeah.

But it's an absolutely you know incredible technology. Yeah.

Um

the way we're approaching it is to say we're going to use logic

and symbolic representations such as code.

Yeah.

To encode the interactivity and deterministic part of the world. And then we're going to use pixel prior to then render the world.

Sure. But the pixel part is only responsible for the appearance.

But games or real world physics, it's there's a lot of deterministic part that is better represented through code and symbolic represations.

Yeah. So I when I when I watched the demo was very impressed. It felt like uh the first like sort of turning an a coding agent loose on like the Unreal Engine platform. Uh are you using a game engine under the hood? Uh, and then how much of a harness did you have to build to actually get this result? Because the the little interactions around there's music playing and you can bowl and you actually have a game within a game. You can go up to an arcade game and play Space Invaders that it was all built. Uh, how much of this is the harness? How much of this do you get for free by sitting on top of Frontier LLMs? And then how what are you using in terms of game engine and stuff off the shelf?

Yeah, great question. So we are we actually are using game engine but we forced an open source game engine and make a lot of customizations on top of it.

Oh probably GDO

our model exactly yeah um to allow our model to essentially leverage a lot of things that offtheshelf models can't.

Awesome.

So

and then so you can think of it as like this code generation model that is we we postrained to have

to be able to use a variety of tools.

Yeah.

To build the world.

Very cool. So, uh, do you want to become a game developer and use your own tool to generate the next super deep world at way lower cost? Do you want this

playing in a game right now?

Yes. the we'll get into simulation theory at the next interview, but uh uh or or or do you want this to be like a consumer tool similar to what we've seen with Sunno midjourney where there are people that go and sort of generate their own images or music and enjoy those for themselves. Maybe they share them, but maybe they just enjoy them for themselves. How do you see this being adopted?

Yeah, frankly on the our ambition is beyond gaming. So, we really want to solve multimodal reasoning. Yeah.

Like be able to allow models to really do long horizon planning. Sure. both the virtual world and the physical world.

Yeah.

But in terms of commercialization, short-term commercialization, we want to empower um empower basically people to be able to monetize with their ideas that is not bounded by their skill.

Yeah.

Right. We want to be the enablement layer to shift leverage from skill or domain knowledge to really taste.

Yeah.

So that anybody with good ideas can then monetize in their ideas.

Yeah. So, uh, how much of how much of a network is important here? Like I'm thinking about the priors of Roblox, building games within the game engine. Uh, and a lot of that comes from the the network effect, the the multiplayer nature. Uh, do you think that will be important as a growth engine for you?

Absolutely. In fact, in our today, you can say, I want to build an open like multiplayer game. Sure. And then the modeling agent would automatically configure

Yeah. database and multiplayer setups for you such that you can actually one click deploy this experience.

Yes.

But importantly that would not be that sticky of a network effect for you. And so what I'm thinking is like do I wind up with a moon lake handle at some point that I can take across the different uh games that are created through the platform?

Yes. Okay. So you will be able to say you know create create an ID and then create 10 10 of your games just monetize on top of less platform and we help you distribute it to the players and help you monetize.

That's amazing. This is very very cool. Uh take us through the funding news. How much have you raised so far?

We've raised 28 million total in

who's in any good

I think that was I think this is the same round as as as last time. But it's it's been it's it's an honor to hit it again.

Yes, I'll hit the gong for anything that Jeff Dean's ripping checks into. Congratulations. It's a great lineup of investors. Uh and thank you so much for taking the time to come on the show.

Yeah, great to get the update. Really impressive progress.

Congrats. We'll talk to you soon. Have a good rest of your day. Let me tell you about Apploving. Profitable advertising made easy with Axon.ai. Get access to over 1 billion daily active users and grow your business today. And let me also tell you about graphite code review for the age of AI. Graphite helps teams on GitHub ship higher quality software