Teaching LLMs how to solid model

Original Article ↗ Hacker News Discussion ↗

Current capabilities and use cases

  • Multiple commenters report success using LLMs (via OpenSCAD, CadQuery, etc.) to generate simple parts: enclosures, ornaments, screw caps, brackets, even a “Terraforming Mars” city toy.
  • Typical workflow: user describes the part, model outputs code, user renders, prints, and iterates; sometimes a script feeds rendered images back to the model for correction.
  • Vision models that accept screenshots/renders can sometimes debug geometry, supporting hopes for RL-style feedback loops.
  • For hobby 3D printing and “good-enough” parts, many find this surprisingly useful; others say it’s interesting but not yet a net time saver.

Model and tooling comparisons

  • Reports that only top-tier models are usable; even then, they frequently mis-handle rotations, coordinates, or mating conditions, especially in complex assemblies.
  • One comment claims a particular model “one-shots” more elaborate parts; others describe OpenAI/DeepSeek as weaker for OpenSCAD and praise Gemini’s visual feedback. Claude is mentioned but not broadly tested.
  • OpenSCAD is seen as a good testbed but limited; alternatives like Build123d, CadQuery, Solvespace, and Lua-based frontends are discussed.
  • A commercial “AI CAD” backend is criticized for weak results; its author responds that the underlying CAD kernel and training data are still immature.

Interface: text vs drawing vs VR vs hybrid

  • Strong skepticism that long text prompts are an efficient primary CAD interface; many prefer sketching, 2D drawings, or direct manipulation (mouse, tablet, VR).
  • Several argue the real future is hybrid:
    • AI suggests parameterized primitives or assemblies;
    • user clicks, drags, or sketches constraints;
    • AI refactors, propagates edits, or bulk-edits features (“change all #6-32 to M3”).
  • Voice + pointing, XR “hand sculpting,” and napkin-sketch-to-model are seen as promising modalities, especially for non-CAD experts.

Suitability for professional/mechanical design

  • Mechanical engineers stress that real work involves tight tolerances, draft, moldability, fasteners, threads, ribs, FEA, manufacturability, and integration with existing parts—none of which current LLMs handle reliably.
  • Many see near-term value in helper roles: generating footprints from datasheets, converting PDFs or legacy CAD, critiquing designs, exploring options that are then verified by simulation.
  • High-stakes structural/mechanical decisions by LLMs are widely viewed as unsafe without strong automated validation.

Future directions and concerns

  • Proposed directions:
    • Domain-specific CAD DSLs that operate on features and queries instead of raw coordinates.
    • Tokenizing breps/3D geometry for transformer-style models.
    • RL training using SCAD libraries and render-based volumetric comparisons.
  • Concerns raised about:
    • AI-generated models gaming financial incentives on 3D model sites.
    • Hype around “AI CAD” vs today’s reality that even mainstream CAD/ECAD tools are slow, buggy, and frustrating.