Why AI can render a perfect Mario but fails at a simple vacuum

It is a bizarre paradox of modern generative AI: you can prompt a model to create a highly detailed figurine or a vibrant cartoon character like Super Mario, yet the moment you ask for a functional wedge ramp to help a robot vacuum climb a doorstep, the system falls apart.

I recently hit this wall personally. I own a Bambu P1S, but I don't have the technical background to sit down and master complex CAD software. I attempted the "describe it and get a model" approach using various text-to-3D AIs, but the results were essentially useless. The geometry was messy, impossible to adjust, and never actually matched the dimensions I needed. I even tried a more advanced AI workflow where I had an LLM agent write Python scripts to build the geometry directly, but it hit a ceiling almost immediately, only capable of generating basic primitives like cubes and spheres.

The breakthrough didn't come from a better prompt, but from a complete shift in methodology: geometric decomposition.

Instead of asking for the whole object at once, I started breaking a complex part into ordered, grouped steps. I describe each tiny segment as a specific technical specification and then let an agent execute those steps within Blender using blender-mcp. This effectively turns the problem into a small engine. The core insight is that we are converting 3D spatial reasoning—something LLMs are notoriously bad at—into structured, sequential code, which is where they actually excel.

I’ve documented this specific approach and the underlying logic here:

https://github.com/zhuchaokn/spec-3d-model

This experience leaves me with two major questions regarding the current state of AI development and prompt engineering for physical objects:

The gap between aesthetics and utility

Why is "functional part" generation so much weaker than "aesthetic/figurine" generation? I suspect it's a combination of three factors:

• Data Scarcity: Most training sets are composed of meshes from art platforms, not parameterized CAD datasets used by engineers.
• Representation Issues: Generative models often work with meshes (which are just a cloud of triangles), whereas functional engineering relies on B-rep (Boundary Representation) to define exact mathematical surfaces.
• The Evaluation Problem: There is no standard benchmark for "Is this watertight?" or "Will this actually print?" Unlike image generation, where a human can instantly spot a bad face, 3D utility is hard to automate via loss functions.

Is code-based modeling the ultimate path?

Is "turning 3D modeling into code for an LLM" the correct framing for the future of generative CAD, or is there a more efficient way to bridge the gap between a text prompt and a physical, manifold object? It feels like we are moving away from "generating shapes" and toward "generating instructions for construction," but I'm curious if anyone has found a more direct way to handle spatial reasoning without the heavy lifting of Python scripting.