I've been puzzled by something: AI generation can produce an elaborate
figurine, a cartoon character, even a convincing Super Mario — yet it
can't reliably make a simple wedge ramp so my robot vacuum can climb a
step. For context: I bought a Bambu P2S but can't model. I tried the "describe
it and get a model" AIs — the output is unusable, you can't adjust it,
it's never quite what I meant. I tried having an agent write Python to
build geometry directly — it tops out at simple primitives.
What finally worked: geometric decomposition. I break a complex part into
ordered, grouped steps, describe each as a small spec, and let an agent
execute them in Blender (via blender-mcp). That process turned out to
abstract into a small engine — the key insight being it converts the 3D
spatial reasoning LLMs are bad at, into the structured code they're good
at. I wrote it up here: https://github.com/zhuchaokn/spec-3d-model
My questions:
- Why is "functional part" generation so much weaker than
"figurine/aesthetic" generation? Is it data (no parametrized-CAD training
sets), representation (mesh vs B-rep), or evaluation (nobody benchmarks
"does it print / is it watertight")?
- Is "turn 3D modeling into code for an LLM" the right framing, or am I
missing something better?