There's a moment you hit when working on AI where things stop feeling impressive.
For me, it came from thinking about what it would actually mean for an AI system to operate in a world.
Not generate one. Not describe one. Not animate one.
Operate in it. Understand it. Live in it.
The frustration
I was drawn to this idea early, the idea that AI could design physical systems, simulate new technologies, or generate entire interactive worlds where the rules themselves could change.
A system where you could say:
- ✦change gravity
- ✦alter thermodynamics
- ✦build a new kind of machine
- ✦construct a world that actually behaves like one
And it wouldn't just render it. It would understand it.
But the deeper I looked into how people were approaching this, the more something didn't sit right.
A lot of the momentum around “world models” today is built on video.
You generate the next frame. Then the next. Then the next.
It looks convincing. Sometimes incredibly so.
But it's not a world.
It's closer to a dream.
If you walk out of your house and come back, it shouldn't eventually be a different house.
But that's exactly what these systems allow. There's no true persistence, no underlying structure enforcing consistency. Just a continuous hallucination that happens to look coherent locally.
That was the break for me.
What most people are getting wrong
Most modern work on world models inherits its assumptions from robotics.
And that made sense.
A world model, historically, is: given an action, predict the next state.
For a robot, that's enough.
The world only needs to exist as long as it's being observed.
It doesn't need to persist beyond that.
But that's not how humans experience reality.
We rely on persistence:
- ✦objects exist when we're not looking
- ✦spaces maintain relationships over time
- ✦environments carry meaning, not just appearance
Without that, you don't have a world.
You have something much closer to a sequence of impressions.
That distinction matters more than it seems.
Because persistence is what allows:
- ✦learning over long horizons
- ✦emotional attachment
- ✦spatial reasoning beyond immediate perception
- ✦actual understanding of systems, not just reactions
So the problem isn't that current approaches don't work.
It's that they're solving a different problem.
The nuance
To be clear, video-based models aren't useless.
They're actually very good at something important: learning cause and effect.
If I open a drawer, what should happen? If I push an object, how should it move?
By compressing the world into visual transitions, these models learn patterns of interaction very efficiently.
That abstraction is valuable.
But causality alone doesn't give you a world.
It gives you behavior.
What's missing is the structure that behavior operates within.
The bet
The bet behind STAR Labs is simple:
Implicit, dream-like world models are not enough. Explicit, persistent worlds are necessary.
Not as a layer on top. Not as an afterthought. As the foundation.
We believe a real world model needs:
- ✦explicit geometry
- ✦persistent state
- ✦consistent rules (what we think of as “game mechanics”)
- ✦the ability to evolve over time without collapsing into incoherence
In other words, it shouldn't just look like a world.
It should hold together like one.
What we're building
Our first step toward that is a native world model engine.
Not a video generator. Not a simulation layer bolted onto an existing game engine.
But a system designed from the ground up to:
- ✦represent space
- ✦encode physics
- ✦maintain persistence
- ✦and support interaction over time
At the core of this is a question most people skip:
What is the right representation for a world?
Right now, we're exploring a unified substrate built from:
- ✦physics-imbued Gaussian representations
- ✦coupled with mesh-based structure
Early results show:
- ✦stable elastic behavior
- ✦promising extensions into fluids
- ✦a path toward handling rigid bodies within the same framework
If this holds, it means one system can represent geometry, material properties, and dynamics all at once.
That's the direction.
Why us
We're not coming into this with 20 years of attachment to a specific paradigm.
That's intentional.
A lot of this field is being pushed forward by people who are finally able to scale ideas they've believed in for decades.
That's powerful.
But it also creates blind spots.
We're approaching this from the opposite angle:
- ✦no legacy assumptions
- ✦no commitment to existing frameworks
- ✦just a focus on what a world actually needs to be coherent
And a willingness to question whether the current trajectory is pointing at the right problem at all.
What this unlocks
If we get this right, the applications aren't incremental.
They're foundational.
- ✦Generative design systems that understand physics, not just shape
- ✦Educational environments that can be created and modified in real time
- ✦Interactive worlds that don't reset every time you look away
- ✦AI systems that don't just react to environments, but actually understand them
This isn't about better rendering.
It's about building systems that can reason within reality, even when that reality is generated.
The direction
We're still early.
Right now, this is about getting the representation right.
Because everything else depends on it.
But the goal is clear:
Build AI systems that don't just generate worlds, but can actually exist in them.