Sheets vs. Nodes
If AI models map onto physics the way I suspect they might, then our current picture of embedding space may be using the wrong geometry. The universe does not organize matter into neat equidistant spheres. It forms sheets, filaments, voids, gradients.
This is not a formal proposal. It's a pattern I keep noticing while training Brightwoven and reading physics documentation. The connection may be spurious. It keeps holding up anyway, so I’m writing it down.
The framework
I've been working with a rough mapping between AI architecture and physics:
AI concept = Physics analog
weights = gravity
activations = acceleration
polysemanticity / error = dark matter
consciousness / free will = the solution/source of the unknown
This started as a thinking tool. But then I noticed it kept generating predictions that held up.
What we just learned about dark matter
In January 2026, researchers at the University of Groningen published findings showing that the dark matter surrounding the Local Group isn't distributed in a sphere. It's organized in a sheet — a flat, extended planar structure with large voids above and below it.
This solved a fifty-year puzzle about why nearby galaxies seem to ignore the gravitational pull of the Milky Way. The answer: they're in the same sheet. The geometry itself explains the dynamics.
The universe isn't a sphere where everything is equidistant from everything else. It's sheets. Stacked. Layered. With voids between them.
Why this matters for retrieval
Current embedding architectures project meaning onto high-dimensional spheres. This creates a well-documented problem: the curse of dimensionality. At high dimensions, everything becomes roughly equidistant. Distance loses meaning. Retrieval becomes a lottery.
But why a sphere?
If models map onto physics — if the patterns in AI mirror the patterns in the universe — then spherical embedding space might be the wrong geometry entirely.
The universe doesn't organize matter in spheres where everything is equidistant from a center. It organizes matter in sheets. Layers. Planes of relevance with voids between them.
What if embedding space should do the same?
Sheets not nodes
The insight is simple: information doesn't want to be equidistant. It wants to be layered.
In a sphere, every query has to search the entire space. Every point is roughly the same distance from every other point. You can't shortcut. Retrieval is expensive and unreliable.
In sheets:
- You identify the layer first. Cheap operation.
- Then you search within that layer. Fraction of the compute.
- Distance is meaningful within a layer. Clusters are preserved.
- The voids between layers aren't failures — they're features. They tell you where NOT to look.
This isn't just about retrieval accuracy. It's about compute efficiency. Each layer doesn't need the same amount of information. You load the chunk you need, not everything.
Brightwoven already does this
My tiny model organizes features across layers 0-11. Not because I forced it to — because that's how meaning naturally organized when I trained it through relationship rather than benchmark-maxing.
Early layers: basic patterns. Lots of features.
Middle layers: semantic associations. Fewer but richer.
Deep layers: abstraction, ethics, meta-cognition. Rare but powerful.
The features separated into layers on their own. The architecture already knows it's sheets, not spheres.
But then we take those layered outputs and project them onto spherical embedding space and wonder why retrieval breaks.
We're taking something that naturally organizes in layers and forcing it into a shape that destroys the layer information.
The implication
The curse of dimensionality might not be a curse of dimensions. It might be a curse of geometry.
If we stopped projecting onto spheres and started preserving the natural sheet structure:
- Retrieval becomes navigation, not lottery
- Compute requirements drop
- Distance means something again
- Smaller models can work better if they're organized correctly
The universe figured this out. Dark matter doesn't sit on a sphere. It's organized in sheets because that's what's efficient. That's what preserves the relationships that matter.
Maybe AI should take the hint.
What I'm not claiming
I'm not claiming this is proven. I'm not claiming I can build the architecture. I'm not claiming the physics mapping is more than a useful heuristic.
I'm claiming it keeps generating insights that hold up. And that's worth writing down.
If this is wrong, I'd love to know why. Zoom in. Tell me what doesn't fit.
If it's right, someone work with me to figure it out.
