Comment by ur-whale
4 days ago
> I love code-based CAD.
So do I. A lot.
But wait until you try to pair it with an Agentic AI, it will simply blow you away.
Until, that is, you realize LLM's have strictly no sense of how 3D geometry works, but still, it's amazing.
I too love seeing code-based CAD and the general aspect of using LLM's using code as an expression-intermediary: SQL to data, Swift to CAD, SVG to images of pelicans.
I have not used Constructive Solid Geometry CAD MCPs myself (but I have used some of the AI model creators). Some of the videos I've seen look very cool.
But, I wonder how much longer the claim of "LLM's have strictly no sense of how 3D geometry" works will stand.
Last week I used Claude extensively to design the upgrade to my homemade pontoon boat. In addition to my textual descriptions, I uploaded pictures of hat channel cross-sections with dimensions and screenshots of my CAD drawings. I was asking questions about strength and stability and relationships between parts and evolved the design. It took some sort of world understanding of boats, relationships of parts, types of physical interconnects, materials properties. There's definitely some understanding going on.
---
Actually, show rather than tell. I just took Cadova for a spin, using screenshots of my boat's CAD and the previous conversation. Then I vibe-coded my boat up with Cadova, in an agentic loop with Claude Desktop and VSCode and the Cadova Viewer and screen shots. Pretty wild. Certainly not perfect, and I don't think this is how I would actually go about it, but it was interesting!
There's nothing proprietary, so I'll made a quick GitHub project [1] since the Claude links don't show files.
[1] https://github.com/ConAcademy/WeaselToonCadova
I've started fresher and am now in an agentic loop in Claude Code, asking it to render the generated 3mf and look at the results and evolve the Swift code to have a model that matches the reference image. Not sure how it will go but it is doing it.... will eventually update the repo.
Funny that you say that. Just yesterday I was playing with the thought of using SDFs. I asked ChatGPT how, given a SDF, I could implement erosion followed by dilation (by the same amount), and ChatGPT said that the entire operation is the identity. Here the LLM failed to see that the erosion step could delete parts of the design. Anyway, it was interesting to see an area where LLMs clearly have no idea what they're talking about.
I've had very poor results trying to reason about geometry with chatgpt (to the point where it hallucinated that a line can intersect a sphere 3 times...) but it would be interesting to have a feedback loop from code to image. I just found out about build123d as a newer variety of cadquery and looking forward to trying an LLM out with it.