[hoakiet98]

This is super interesting! Cursor CEO mentioned in interviews that they initially started with building AI for 3D models, but pivoted because they couldn't get enough data for the models to be effective.

I wonder if you think this is still true given how much better the foundation models are now.

[abe_m]

In all the commercial mechanical CAD packages I've seen, the default libraries are pretty weak. Each organization needs to dig in to build out the standard parts libraries to cover the basics of fasteners, bearings, etc. I think the data needed for CAD AI to work properly is going to be processing the various industrial standards from SAE, DIN, ISO, AGMA, etc to properly implement standard interfaces. I suspect it is going to be more of a tool matching AI, where there are library calls that have the explicit technical definitions of various mechanical items, and the AI is coordinating them.

But a lot of the data needed to train an AI is probably locked up in user company product lifecycle management systems. Autodesk is probably going to be out front in any AI wave of CAD with all the data they are accumulating in their Fusion 360 cloud. The main question for me is whether they will pull a Kodak and not develop it for fear of cannibalizing their current market, or not.

[zachdive]

yes i saw that! imho the latest foundation models have enabled a whole host of new possibilities. particularly improvements on SWE-BENCH and other software related benchmarks seem to translate fairly well to openSCAD as we're generating code. however there is still a lot of work to do, as these models struggle to reason spatially. gemini 2.5 is probably the best here rn.

the future of cad generation will undoubtedly extend far beyond simple code-gen to generating human-readable features. this is where the data will be important!