I’m still learning about prompt-driven AI development for CAD, and I’m curious what people here expect to see over the next year. Do you think it will become useful for creating editable models from prompts, or will it mainly help with concepts and simpler tasks?
Are you experimenting with emerging CAD platforms, such as ZooCAD, or following AI developments in established systems like SolidWorks and NX?
TIA for any replies.
IMHO, ‘AI’ that’s bolted on existing CAD systems looks like more of a case of FOMO then addressing the actual issues.
Could it be helpful? Sure, as any tool would be, for specific use cases, but it’s kind of funny that pretty much no one talks about its limitations, which are aplenty.
IIRC, Autocad trained their LLM model on a couple(?) of million CAD models.
How they plan to use it in F360, I have no idea (or really care, TBH :))
Some people do use LLM’s to create models in OpenSCAD, which makes sense, since that’s close to how modelling is done in OpenSCAD.
So far, all prompt->3d model CAD startups struggle with even the simplest models, and produce
garbage output. Will that change? Who knows.
Are those models ‘good enough’ for exploratory/design phase? Maybe?
Is using them faster and economically cheaper then trained and competent human doing it from scratch?
No matter how good/advanced ‘AI’ gets, there is still need for someone with actual expertise to guide it, so it doesn’t create costly mistakes.
Anecdote I heard is that some company, enamoured with the ‘AI’ promise from the vendor, fired their CFD/heat transfer experts (‘AI can do it cheaper’).
The result? 40+M$ manufacturing loss, because they had to recall their product since it tended to overheat in use and either randomly work or die altogether.
A year is a long time in this game, but I think we will see where the LLM’s fit in CAD when the cycle is past the hype phase.
This wall of text doesn’t really answer your questions, I know 
It’ll looks like AutoLISP.
It’ll take longer way more typing to get anything done.
Just like post in Linkedin.
Someone used AI on SolidWorks modeling challenge and claim AI beat the human time.
They didn’t include the 2 days they used to train AI.
It’ll take a long time to train AI not to hallucinate to be useful in CAD.
I would rather see the focus used to automate tedious application process steps and problem diagnosis and repair first. Broken link… AI find it. Broken assembly relationship or feature tree… AI fix it. Imported STL to reverse engineer… AI make it an editable Brep.
Reduce the cognitive load of the CAD user so they can focus on design instead of the operational tedium of the CAD application that distracts from design.
Couple of years ago we sub-contracted some CFD to a company using using AI software to design some better heatsinks for us. The software kept blocking off ~70% of the fan outlets and never got better than just regular heatsink pin geometry. So we binned that off. Hoping it’s improved since then, the software vendor were also confused by its output but we didn’t have the project time to wait for them to troubleshoot.
Yes, stop this “it will design for you and create all your drawings, no engineer required”.
Just help me find stuff in a huge assembly tree. Select all fasteners with a prompt, or all parts smaller than a certain size, show only sheet metal parts, etc. Create 10 configurations with dimension “L@Sketch1” set to these values in this spreadsheet, then suppress “cut extrude 8” in the first 5 configs and rename it “Mounting Holes”.
So much low hanging fruit users would be excited for, it just isn’t as flashy as watching it create a poorly thought out design, in addition to being poorly modeled. Looks good to people who don’t know any better.