News for hackers. Stuff that matters.
Re:How to CAD Almost Anything by eadz on Friday November 20, 2026 @01:06AM (#50570770)
I love open source ( although slightly less relevant now ), and so I spent many many hours using OpenSCAD to write parts that in hindsight would have taken me 1/10 time in something like Fusion. When I finally learned Fusion, I regretted not learning it earlier.

However now everything has changed. I personally believe that the days of designing by hand are numbered. I think the current apps are a bit like the early days of AI assisted coding - autocompleting certain manual steps. Compared with full agentic "solve x" prompts.

I've been working on my own AI-First CAD ( and CAM! ), thinking about everything from first principles for AI - is it easier to think in terms of operations? or start with the finished product? I'm pretty happy with the engine ( and kernel ) I've (+agents) built, it can do full assemblies, including interactive sliders to move them etc, and then do the 3 axis CAM ( or export to STL for printing). I'm really interested in genuine feedback if anyone has a few minutes to try it out www.zelm.ai.

Re:How to CAD Almost Anything by adamweld on Friday November 20, 2026 @01:54AM (#50571052)
As a mechanical and software engineer I've been following this space closely, and similarly thinking about what interfaces will power the CAD packages of the future.

I checked out your project but was a bit disappointed to see that it's closed source. As you're probably aware in the last 6 months there have been a ton of new OSS B-Rep kernels and CAD programs hitting the ground, many of them working towards the same goals that you and I have.

It's my personal opinion is that in this new age of AI coding, the best way for a project like this to gain momentum and user base is to make it free and open for other users/contributors to work on and with.

Just my 2c.

Re:How to CAD Almost Anything by eadz on Friday November 20, 2026 @02:02AM (#50571110)
I totally get the sentiment, but I think software has changed quite a bit. If we're honest, AI 99.9% built this, I just provided the vision.

I've contributed to open source, and ran my own open source projects in the past, but when everyone can have the software exactly they want it's just a different ballgame. I think the value of Zelm isn't so much the code, but the infrastructure around it to make it a seamless experience.

If there is a feature you want, just ask for it in the chat or via MCP, and it will be added to the backlog automatically.

Open source was never about free as in beer, but in this case Zelm is pretty close to free if you use it as a MCP server.

Re:How to CAD Almost Anything by adamweld on Friday November 20, 2026 @02:41AM (#50571404)
I've been having some interesting discussions with friends about this exact topic, whether AI will be a boon to OSS, or the end of it.

I think that many tools, services, and programs will be more personalized, and maybe less shared. But that the core building blocks of the tools will stay open and continue to be refined by the community.

When I'm coding with AI, one of the main reasons it's so capable is the ability to build off of languages, packages, and frameworks that are battle tested across many deployments. The end result is better if my (+ my agents) focus is on what's new and different about what I want to create, not reinventing the wheel. So I'll import a proven serial library rather than writing my own, pick the right programming language for my use case rather than writing in assembly.

With regards to your project and CAD, I think it's a clean and efficient interface, and part I generated looked good. Personally I need more traditional CAD features for my day to day (drawings, multi part assemblies, hand editing) but it looks like a good tool for the right use cases.

To finish my open vs closed source thought - have you considered open sourcing your b-rep kernel? Your app may be the secret sauce, but the kernel is boring but vital core. Parasolid has been developed over the last 35 years and is the reference point. NVIDIA open sourced SMLib but it is buggy and incomplete. OCCT exists but doesn't handle fillets well. Zoo spent two years with tens of developers building their kernel from the ground up (also unfortunately closed). There are a dozen new kernels written in Rust on github but none are complete.

I really think that if all of the smart people trying to roll their own kernel worked together and put effort into really completing a good open source kernel, it would be the start of a cambrian explosion of CAD tools and AI enabled hardware engineering tools. Food for thought.

Re:How to CAD Almost Anything by eadz on Friday November 20, 2026 @03:03AM (#50571553)
Have you seen https://solidcraft.io/solidcore/ ? It's a rust CAD kernel and open source. It's a port from SMLib. It also seems to have an active open source community.

I don't really know much about CAD kernels, and so I really don't think I should be running a cad kernel open source project/community ( which is what successful open source is about ) I wouldn't be able to decide if we should do feature x or not.

> Personally I need more traditional CAD features for my day to day (drawings, multi part assemblies, hand editing)

Zelm supports drawings ( including custom drawings, see Download button ), assemblies ( with movement interfaces, click on Features ), and while it doesn't support hand editing ( i.e. with a mouse ), designs are parametric so you can tweak them ( click on Params down the bottom ).

This is an example: https://www.zelm.ai/p/v8nzvHBhz7?share=Kis1TuAUtwn7nR3VNx-rj...

Re:How to CAD Almost Anything by imtringued on Friday November 20, 2026 @03:24AM (#50571739)
>I don't really know much about CAD kernels, and so I really don't think I should be running a cad kernel open source project/community ( which is what successful open source is about ) I wouldn't be able to decide if we should do feature x or not.

Is this some kind of joke I'm not getting? People use AI to do things so far outside their competency to the point where they have no idea about anything? I mean, who exactly would trust such a project? I'd rather do it myself.

Re:How to CAD Almost Anything by eadz on Friday November 20, 2026 @03:38AM (#50571845)
> People use AI to do things so far outside their competency

yes, if you want to do things far outside of your competency

> I mean, who exactly would trust such a project?

Depends how it can be verified. In this case there are potentially easy solutions, print it on a 3d printer and measure it. Is it 30mm across?

Re:How to CAD Almost Anything by reillyrider on Friday November 20, 2026 @03:07AM (#50571589)
I'm not really sure about AI CAD. The things I CAD are very specific, I can't just go "give me a generic box", it needs to be these exact dimensions, it needs to have cutouts at this position with this size, it needs holes at this spot, manufactured with these materials and tolerances, then I need this all nicely parametric and easy to change. If I need to put that many specific details in my prompt (and then double check the work afterwards) I might as well just do it myself.

And how would iteration look like? It's a million times faster and easier to just navigate around a 3d model, click on the feature to edit, start editing. Describing which feature for the AI to change seems like a nightmare.

Re:How to CAD Almost Anything by WillAdams on Friday November 20, 2026 @08:46AM (#50574589)
I tried to use an LLM for reasoning about a quite simple 3D project (a box with rabbet joinery) and it created a convincing UI and so forth, but when the DXF was examined, it was _not_ workable.

I was able to (eventually) get a working file by laboriously describing the geometry of each part and its joinery in exhaustive detail:

https://community.carbide3d.com/t/a-simple-cnc-box-in-depth/...

Next up is adding a hinged lid to the existing sliding and no-lid options.

Re:How to CAD Almost Anything by kabes on Friday November 20, 2026 @05:50AM (#50572892)
Thing is, I find natural language a really bad fit it describe a 3d model.