I just stumbled across https://illoca.com/ which is the best AI CAD that I’ve seen so far. It still seems to be missing the answer to “OK the AI made a thing, but now I want to interact with the resulting file manually”, but it does seem to be able to make decent corrections to model. Using the free credits they give out we made a pretty decent plywood flat pack table in two prompts:
5 Likes
Yes, and…
I’ve been really enjoying OnShape’s Feature Script coding with Codex. I can use natural language to accomplish most basic construction.
I’ve also been using the “full duplex voice mode” to record measurements no-handed. The other day I walked into my shop, opened the conversation mode, asked it to help me take some measurements to help identify the dust port adapters I should print for my new dust collector.
Then I walked around the shop saying “Rikon sander dust output outer diameter 1.75in”, or “Dwalt Chopsaw outer diameter 2inches, 1/8th inch thick”. You can ask it to read the records back, so it would be like “Got it, Rikon sander, 1.75 inches”.
Then I took that info back to Codex, and said (literally):
“[link to the voice chat measurements thread] I made a new onshape doc [link] can you get it going with the resolved adapters dimentions from this linked chat? Wall thickness should be a Configuration, as well as the two diameters, so this could make any tube coupler sizes in this design. Design is, imho, a simple rotated solid, with a shell operation along the through axis. Sketch, from origin, draw the radius/cross section of the adaptor with a smooth drawing tool, then rotate the design, and shell it. Yeah?”
3 Likes
We are living in the future 
Cool!
I’ve been experimenting with Codex and OpenSCAD, building parts step by step, I’m missing being able to tune parameters in a GUI (like you have going on there with OnShape and Illoca) but I do have them. I seem to be able to get quite complex functional designs printed and a common design language naturally builds up around them (ready for the next project).
I have found limitations with the model’s spatial awareness, and that you can burn a lot of tokens trying to close that gap when actually your feed back and getting them to annotate their drawings can do more.
I’m finding it fascinating and frustrating! What else are you using for this kind of thing?