Design for 3D-Printing

This looks so good. I’ve gotten into 3D printing in the past six months with an A1 Mini. I initially bought it intending solely to do creative projects with my kid, but I’ve been surprised to find myself getting deeper into printing functional parts. I recently printed a 6” server rack for a GLi.net Beryl and Apple TV for travel, from a combination of pre-designed and self-designed parts.

3D printing as a pursuit can be time-consuming - there’s always a risk with these things that you take them on as a dilettante and they end up gathering dust in a corner. I initially scraped by with some middling Blender skills (leaning into non-destructive operations where possible), but that is far from ideal - you really do need CAD. But to anyone considering jumping in, I would say: if you get an A1 (get the full size, not the Mini) and use Claude to write your parametric OpenSCAD scripts, the time commitment is such that you can _just about_ indulge in this hobby as a dilettante - eg, as a project for your kids. Without LLMs, I think it would be too much of a commitment unless you’re really dedicated, or already have CAD skills.

Anyway, gonna go read this in full.

the Bambu P1S with AMS is one of the better purchases I've made. I've had an ender3v2 for so long and while it worked ok (and arguably better than many people's experiences with them), I got tired of constantly fiddling with stuff.

Now, it just works. It doesn't matter what I throw at it. Made me get into the CAD hobby too.

>use Claude to write your parametric OpenSCAD scripts

Can you talk a little about it?

If you can assemble Legos you can design in Tinkercad. You don't need to mess around trying to get LLMs to write scad files, though the results can be hilarious.

Side bar... There are a lot of people who are going to use LLMs to try to do 3D modeling stuff and who are going to hit a wall with it really, really fast. Mechanical design really is a completely different discipline that is very poorly abstractable in the particular way that software engineers are used to.

Worth noting: TinkerCAD is capable of parametric modeling in "CodeBlocks" mode. I prefer writing OpenSCAD (by hand, rather than via a warmed-over markov chain), but having the option within something as inviting as TinkerCAD seems great for beginners.

I use ChatGPT to help me with OpenSCAD and really enjoy it. It doesn't always nail it, but often works as a shortcut, especially with loops to create radial extrusions. You can muck around with things like asking it to give you code to design a castle, and then see how off-track it is.

I disagree. My co-worker is an industrial designer and uses Rhino as his day to day CAD tool of choice. I was delighted to see that it translates everything you do to a command line syntax and there's also Python integration. We did some simple tests the other week and you can actually prototype reasonably well by instructing an LLM to generate the Python code that creates the models. It still requires fine tuning but seems like a similar multiplier like using LLMs for programming to get the boring boilerplate out of the way.

I'd say a good designer will at least 2x, probably 5x. We are preparing to test with students next semester to see how non-experts profit from this.

Edit: I realize I was imprecise, I specifically was referring to 3D CAD for mechanical parts of the type you would generally produce a print for i.e. with controlled geometry and dimensioning.

A) That's a surface modeler.

B) Parametric CAD doesn't have boilerplate in the same way as software. In a part, you have a feature tree, and a lot of thought goes into constructing the feature tree in a way that both allows for reconfiguration and also somewhat resembles the manufacturing process. Every step depends on the previous step in a way that is necessarily impossible to isolate. If you make a step without being aware of the end state of the tree, you will probably have to redo a bunch of the tree if the part is complex.

Also that's not what mechanical engineers spend most of their time doing. Plenty of MEs can code and use an LLM btw, nothing stopping them from optimizing their work processes if they saw fit, yet you don't really see the shoehorning of AI into the space yet.

Exactly. People act like if an LLM can't get it 100% perfect right away then it's a useless tool. Maybe they're be able to do that one day.

But in the meantime I think it's brilliant if it can rough something out, then a human can go and tweak it to the correct dimensions, fix issues or simplify needlessly complex features.

Same thing for artists as well - many will come to realise that sure it generates slop, but that slop is still useful as a skeleton to overlay a human creative vision onto.

I can't tell you how many times I've had to redo my feature tree because I built a feature on a part out of order. Tweaking dimensions that you didn't expect to have to tweak very often also breaks your feature tree.

If you don't know what a feature tree is, it's really worth investigating how parametric CAD is actually done before insisting that LLMs must have a business case for MCAD users!

For hobby widgets and gewgaws you can absolutely make an "approximate object" in Blender, spit out an STL and bounce it to your 3D printer and probably maybe have it work. That workflow is not really used in industry.

Yes, it’s a reasonable comment. Hopefully my original use of the term “functional” was sufficiently qualified by the word “dilettante”, but just in case: no, LLMs won’t turn you into a proper engineer or designer - or even be as good as actual CAD skills - but if it’s server rack bits n bobs, or a Steam Deck dock you’re after, this can work and is a lot of fun. It can also be an on-ramp to learning actual CAD by lowering the activation energy - especially for kids.

Tinkercad seems pretty popular, I definitely looked at it along the way, but ultimately I’m never going to choose to use a product from Autodesk if there’s any reasonable alternative, and I’m usually not going to choose something that’s not open source - especially for a hobby. Different strokes though, might be a good fit for some people.

Yes, as sibling mentions, there’s not much to it: just crack open Claude or ChatGPT and start asking for what you want, as an OpenSCAD script. You’ll quickly get a feel for it. Others might have more structured approaches they can share.

Huh? Claude isn't open-source and neither is Bambu.

No, but those pieces don’t have a learning curve, and I can swap them out easily (Bambu Studio is a fork of Prusa Slicer). If I spend the time learning Tinkercad, then Autodesk enshittifies it, I’ve got to re-skill in another CAD program, don’t I.

And starting your comments with “Huh?” is rude.

I _know_ what a feature tree is, I use fusion every couple weeks. CAD is not some magical hard thing to do, except perhaps for more complex parts where it's a software problem (like mirroring the features of a surface to the other side but being unable to flip them...thus text is mirrored, thanks Fusion). Constraints are indeed a pain in the ass but that's literally their purpose.

And my opinion is still that an LLM could rough out a simple model just fine. It won't be designing engine parts any time soon but simple parts I still reckon it could rough 'em out.