The problem I'm facing is mounting a sensor on a barbell sleeve. The sensor is a semi-hollow 100mm cylinder, 60mm diameter, with a 50mm diameter 50mm deep tube cutout for the sleeve. (It's a bit more complex than that.)
I'm trying to brainstorm different clamping mechanisms, and evaluate durability and manufacturability of several designs.
No matter what I try, the output of all LLMs is absolutely wild, especially if you ask to make a technical drawing. Anthropic reliably been the worst, and Gemini, surprisingly, the best.
But none understand what goes where in any detail, besides spewing paragraph and paragraph of wild ideas including rotational cam clamps over a recessed ring TPU collet (apparently, a "brilliant pivot") and stuff like this.
Is this the "unconquered frontier"? Do I need to blow dust off my AutoCAD? :)
Recommend doing this in a coding harness not a chat box.
The reason I think you might have more success with this is that the model is mostly thinking about the part in words, which it can convert to a part design in CAD in code. LLMs are really good at coding. Also means it can use relative positioning and relationships.
You will be able to iterate more easily, compare things, compute properties, commit to git etc. The process is more reproducible and steerable than generative production of images.
When the LLM can look at renders of the geometry it generated, it’s easier for it to discriminate when it’s producing nonsense like misaligned parts, things that don’t fit, etc. It’s still going to kind of suck, but it will be better. The whole process of code -> render -> inspect forces the model to put up or shut up and provides grounding. Meshes > bloviating.
As far as I know today's LLMs don't have a "visual imagination" but a process like this could be a slow approximation of one. They clearly do have SOME spatial understanding (pelican tests show us that!) but it feels really non-human.
One thing missing from this is kinesthetics. Personally I am mostly not thinking in accurate visuals in mechanical design. I am imagining how the parts feel and kind of how they move and what slips first and what bends and what feels heavy. Imagining what my hands would feel. But I don't think I trust LLMs to evaluate that stuff by writing simulation code yet.
I'm not sure from your description if you control the geometry of the sensor part (i.e, can the clamp be integral to the sensor housing or does it need to be a separate part) also ignores the internals, etc etc.
One thing I noticed is that off the bat, it did think about the assembly in general terms but would need guidance to think harder about FDM limitations and layer orientation etc. These are not good designs for printing. A good dev loop and git history help with these kinds of revisions.
The general principle is that if your domain is verifiable at all, give the model tools and a workflow that constrain and check its output. You want the LLM arguing with the geometry kernel instead of you.
To fully close the loop you print parts as fast and possible and concretely see where they suck.
I'm sure there are ways to "correct the spatial reasoning of LLMs" (whatever that even means) but I'd question whether you even need an LLM to begin with. I know that isn't the answer you're looking for, I'm just a slack-jawed knuckle-dragging Luddite. But this seems like the kind of simple, common problem for which established solutions already existed, like "talk to an actual engineer and not a chatbot that can't even do basic math."
And if it's a matter of not wanting to pay someone, consider that if you're already paying for an inadequate simulation of knowledge, you should be willing to pay for actual knowledge.
I am working on it alone and don't have access or extra money to pay an actual engineer. If I can squeeze 10% of what an actual engineer would've told me, for $20/month I'm in.
And it's not an either-or game, I can have some rough, potentially totally stupid ideas and then validate them with an engineer, instead of paying $10k to make it from scratch.