Ask HN: Honest question about text-to-CAD. Is it BS?
6 comments
My observation is that there's a certain group of programmers who "don't get graphical CAD" (their words) and prefer code based CAD. Now with LLMs, many of them like to spin up their own solution of "text to CAD", "image to CAD" or just another "code based CAD" and then do a "Show HN", praising it as the bee's knees.
When I look at their examples, the objects are usually non-functional, lacking precision and control, and could much more easily be done in a graphical CAD application.
I think some people just don't want to learn graphical CAD and hence are unable to see where their own solution is lacking important features.
When I look at their examples, the objects are usually non-functional, lacking precision and control, and could much more easily be done in a graphical CAD application.
I think some people just don't want to learn graphical CAD and hence are unable to see where their own solution is lacking important features.
I have been doing a lot of related work in AI lately and I would agree that spatial awareness is not its strong point. I even asked it that exact question and it agrees :)
Spatial awareness is the least of the problems... Text is basically middleman which needs to deliver user's design/engineering/imagination to agent. And currently, it compresses a lot.
Amazing, with all that it can do in 2D…
yeah, and now Claude + fusion 360 integration. X is full of bs again. Connecting cad is not the same as doing cad.
I have doubts.
The demos are cherry-picked. Badly. Landing pages, X, Linkedin, all of it.
A lot of these tools just generate meshes and call it CAD. Technically true, but the real value is in BREP. Some are just coding agents that use openscad, cadquery, build123d, etc.
I did actual user research. Talked to people who use FreeCAD/Fusion360/SolidWorks daily, various skill levels (beginners, experts, etc.). Actually paid them actual money to try Text-To-CAD. Almost everyone said the same thing: they're faster in SolidWorks etc. than any Text-to-CAD tool.
Current LLMs SOTA breaks the moment you have multiple bodies or real mechanical parts. Some devs will say they pulled off complex stuff. Sure. But would it have been faster in the GUI?
LLMs are bad at visual reasoning. I don't care what the latest benchmark says. Spatial and geometric reasoning is a weak spot currently.
So is Text-to-CAD just BS right now? Really? Are we chasing the wrong workflow?
Genuinely want to hear your takes to at least to answer myself.