Waveguide/1985
Beyond the Hologram: Why AI "World Models" Need Volume, Not Just Pixels
The AI industry is currently caught in a frantic gold rush to build "world models"—generative simulators designed to act as digital training grounds for robotics and spatial computing. As industry leaders point out, the pitch is seductive: instead of predicting the next word in a sentence like a traditional language model, a world model takes an image or visual state and predicts the next visual frame of reality.
It sounds revolutionary, but beneath the glossy marketing demos, a familiar structural bottleneck remains.
The Hollow Shell Trap
Current approaches to spatial generation rely heavily on surface-level primitives: massive neural networks predicting pixels, stitching together hollow polygon meshes, or throwing millions of Gaussian splats and point clouds at a scene without true physical grounding.
While splats and point clouds are brilliant for capturing the appearance of light bouncing off an exterior surface, they suffer from a fatal flaw: they are completely hollow. They have no concept of internal volume, mass, structural constraints, or physical reality. They are digital ghost towns—skin-deep wrappers painted to look convincing from a single fixed camera angle, but completely blind to the volumetric "bulk" that gives an object its
actual physical presence.
Trying to train a robot or simulate a true interactive environment on hollow pixel-prediction engines is why labs are hitting a thermodynamic wall, forced to burn billions of dollars and haul off-grid gas generators to remote data centers just to keep their brute-force compute running. You cannot out-muscle fundamental physics with
raw surface-level volume.
There is a better way
To build a truly immersive simulation—a real-world-grade environment or a functional holodeck
—one needs the following:
The integration of four distinct 3D mathematical languages
- The Voxel is the "Field" (The Container).
- The CSG is the "Form" (The Construct).
- The B-Rep is the "Outline" (the Topology).
- The Polygon is the "Focus" (The Detail).
The path forward isn't more brute-force pixels; it's a synthesis of structural paradigms:
Voxels for the Bulk: Use high-density volumetric data (voxels or point clouds) to map and define the interior mass—the unrendered "bulk" of an object that gives it structural weight and gravitational presence, much like the unrendered interior of a solid shape.
Illustration of a voxel grid containing color values
The Revolve-to-CSG Pipeline: Instead of letting neural networks guess 3D geometry from ambiguous 2D pixel transitions, sweep or revolve spatial data through coordinate space and snap it into a strict Constructive Solid Geometry (CSG) solver with boundary representation B-Rep to define the topology.
Polygons for the Skin: Reserve clean polygons and high-resolution geometry strictly for the Boundary Representation (B-Rep) interface where that bulk meets the outside world and where detailing is necessary to give definition to the world being depicted.
When you govern a simulation with hard topological and geometric rules rather than probabilistic pixel interpolation, the system stops "hallucinating" what happens when an object interacts with its environment. It calculates the exact
B-Rep boundary intersection natively.
Less is More
True spatial intelligence doesn't come from rendering every empty voxel of empty space; it comes from understanding the clean, underlying source code of reality. By combining volumetric bulk with rigorous boundary constraints, we can bypass the brute-force energy trap and build simulations that don't just look real—they are structurally sound
from the inside out.
The era of hollow pixel-stitching is nearing its expiration date. The future belongs to the geometry.
From 2 to 3 Gemin/Robert E.
.jpg)
(b).jpg)
