Why can AI generate Super Mario but not a wedge ramp for my robot vacuum?
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?
Discussion Highlights (4 comments)
chews
Ive had great success with asking claude to use openscad for part design. Because it’s parametric design as opposed to modeling, it inherently lends itself to making it easier to build. Use metric units when prompting, and Ive even drawn sketches to add to context, and be surprised at how good the outcome can be!
butvacuum
modeling isn't the hard part. It's decomposing it into usable chunks, the knowledge of what's going to be a problem, etc. either way, give an actual CAD program like fusion a go. No, not the FOSS stuff- the real stuff. Although, some of them are "close, but no cigar" these days. Not like gimp which is an exercise in self-flagellation.
pukimak
i asking "why"
YuechenLi
Well, because you need a geometric modeling kernel to generate BRep CAD models, it really has nothing to do with training set as it is very, very, difficult to make one. Lucky for you, I made one that you can try out for vibe-CADing. https://github.com/yuechen-li-dev/Aetheris/ I don't really want to self promote too much here, but currently the other options to do LLM CAD are ecto's vcad, use CadQuery/build123d to access OCCT, or get Onshape for FeatureScript, in case you want to compare the field.