Unlocking the Black Box in Space: Why 3D is the Next Frontier for AI Interpretability

#SpatialComputing #MechanisticInterpretability #WorldModels #AI #3DGeneration

Cross-posted from oourmind.io — part of an ongoing series on the 3D Interpretability Lab.

We have gotten remarkably good at asking what neural networks know. Mechanistic interpretability—the field dedicated to reverse-engineering how AI models work internally—has made massive strides on language models. We can now pinpoint circuits that detect curves, map attention heads that implement induction, and isolate linear subspaces that encode factual associations.

But spatial models—the systems that understand, generate, or reason about 3D environments—remain stubbornly opaque.

This isn't for a lack of curiosity; it is a lack of handles. The internal representations of most vision and world models simply aren't structured in a way that makes them easy to probe, intervene on, or interpret.

That is exactly what makes World Labs' recent essay on "3D as Code" so compelling—and so foundational to the future of 3D interpretability research.

---

📖 The Spatial Lexicon: A Quick Glossary

Before we dive into the architecture, here are the foundational concepts you need to navigate this space:

• Mechanistic Interpretability: Think of this as neuroscience for AI. A subfield of safety/alignment research focused on reverse-engineering how a neural network computes its outputs, not just what it outputs.
• Activation Patching: An intervention technique. You replace a model's internal activations at a specific layer with those from a different input, allowing you to trace which internal computations cause which behaviors.
• Probing: Training a tiny classifier on a model's internal representations to see if a specific concept (e.g., "depth," "surface normal," "object identity") is linearly encoded in the activations.
• NeRF (Neural Radiance Field): An older, famously opaque method for implicitly representing 3D scenes inside a network's weights. You query it with position and viewing direction; it returns color and density. The "scene" lives nowhere you can easily inspect.
• Gaussian Splatting (3DGS): A modern, faster alternative to NeRF. It represents a scene as a cloud of 3D Gaussians (think fuzzy ellipsoids). Crucially, these have explicit parameters: position, orientation, opacity, and color. They are inspectable artifacts.
• Residual Stream: In transformer architectures, this is the vector that flows through the model, additively updated by each layer.
• World Model: An AI that builds an internal representation of an environment to simulate how it changes over time. Vital for robotics, game AI, and spatial reasoning.

---

🏗️ The World Labs Argument: 3D as Code

World Labs makes a bold, structural claim: 3D representations are to spatial AI what code is to software.

The alternative—collapsing everything into a single, end-to-end model that maps inputs directly to raw pixels—is like asking a language model to be the compiled program instead of writing the script. It might work, but you sacrifice the very affordances that make code powerful: inspectability, composability, and reusability.

Their flagship model, Marble, is built entirely around this philosophy. It generates structured 3D outputs rather than raw pixels. Their experimental interface, Chisel, allows users to input coarse 3D layouts (walls, volumes, planes), which Marble then renders into rich, detailed scenes.

---

🔍 Why This Changes the Game for 3D Interpretability

1. Gaussian Splats as Ground Truth Geometry

Most vision models spit out pixels or bounding boxes—outputs devoid of explicit geometric structure. Marble, however, externalizes Gaussian splat parameters as concrete data points.

This unlocks something incredibly rare in interpretability: the ability to correlate internal activations with explicit geometric ground truth. With exported splats, we finally have a tangible reference to probe against. Does the model's residual stream encode splat positions linearly? Do specific attention heads track surface orientation?

2. The Factorized Stack as a Dissection Surface

World Labs advocates for a factorized architecture: separating perception, generation, and rendering into distinct components connected by 3D interfaces.

For researchers, every handoff between these modules is a natural interpretability seam. At every boundary, we can pause and ask: What does this module "know" about 3D structure, and how is that knowledge encoded?

3. Chisel as a Causal Intervention Tool

Chisel—the interface that turns coarse layouts into rich scenes—is essentially a ready-made intervention setup.

In standard activation patching, you modify an internal vector and watch the output shift. With Chisel, you can modify the explicit input geometry (move a wall, resize a volume) and trace how that spatial shift propagates through the model's internal representations. It is behavioral interpretability without needing raw weight access—a spatial version of causal tracing.

4. The Scene Graph Hypothesis

This raises the most theoretically tantalizing question: Does Marble internally maintain something akin to a scene graph?

A true scene graph separates geometric structure (where things are) from appearance (lighting, texture). If the model has learned this factorization internally, we should expect to find:

• An interpretable subspace encoding layout that is mathematically orthogonal to the subspace encoding appearance.
• View-invariant geometry features that persist regardless of camera angle.
• Causal separation: editing geometry activations changes the structure but leaves the style untouched.

Testing this hypothesis would be a clean, novel contribution at the direct intersection of spatial engineering and mechanistic interpretability.

---

🧪 The Research Agenda

For a 3D interpretability lab equipped with access to Marble's weights or API, the roadmap is clear.

---

⏳ The Bottom Line: Why Now?

Three massive shifts are converging at once:

1. Methodology is mature: Mechanistic interpretability tooling (transformers, probing, causal tracing) is finally robust enough to migrate to new domains.
2. The handles exist: World models with explicit 3D structure (like Marble) are newly available, giving researchers the hooks they previously lacked.
3. The stakes are escalating: As world models are deployed in robotics, digital twins, and physical simulation, understanding their internal representations is no longer just an academic curiosity. It is a critical safety requirement.

The World Labs essay frames "3D as Code" as an engineering choice. For interpretability researchers, it is an open invitation.

---

📚 Further Reading

• 3D as Code — World Labs Blog
• Marble World Model — World Labs
• World Labs API Platform
• Circuits — Distill.pub
• Transformer Circuits Thread
• Gaussian Splatting Paper (Kerbl et al., 2023)
• NeRF: Representing Scenes as Neural Radiance Fields

---

This article is part of ongoing research at the 3D Interpretability Lab, developed under oourmind.io. If you're working on spatial interpretability and want to collaborate, drop a comment or reach out.