You know that feeling when an AI spits out a paragraph of mechanical genius, and you catch yourself nodding along—until you realize the clamping mechanism it just described would physically intersect with the sensor mount? Yeah, I’ve been there too.
I’m working on a multi-part 3D-printed product. A sensor sits on a barbell sleeve. Simple geometry: 100mm cylinder, 60mm diameter, with a 50mm deep cutout. I asked the highest-tier reasoning models—Gemini Pro, Fable, K3 Max—to brainstorm clamping mechanisms. Every single one responded with absolute, confident nonsense. Cam clamps over recessed rings. TPU collets. Brilliant pivots that don’t exist in 3D space.
Here’s the uncomfortable truth: The more articulate and confident an LLM’s mechanical suggestions are, the more they mask the absence of true spatial understanding. The most fluent output is often the most dangerously misleading.
This isn’t a prompt engineering problem. It’s not a model size problem. It’s a fundamental architectural limitation. Language models predict text, not geometry. They don’t ‘see’ a 3D space—they weave plausible strings of words that sound like engineering. The more you feed them with reasoning steps, the more elaborate the fiction becomes.
One HN commenter nailed it: ‘Everybody is working on (or waiting for) world models. Language models are not sufficient for doing work in the meatspace.’ We’re all waiting for that mythical ‘world model’ that will finally ground LLMs in physics. But that’s a long wait.
Another comment was even more brutal: ‘It’s wild to me that this is even a question, because of how AI brained the entire world has become. You have a tool that you know works, and a tool that you know doesn’t, but you still feel the need to make the square peg fit into the round hole.’
So what’s the real fix? Stop asking for language. Start forcing geometry.
One clever user suggested: ‘Ask it to model its proposed designs in OpenSCAD or build123d—something queryable. Then render it. Examine plausibility from different angles. Get it to render the part in use and think about forces and motion.’
That’s the twist. The real fix isn’t waiting for a ‘world model’ or improving prompts—it’s forcing the LLM to express designs in executable, queryable geometry, then rendering and testing that output so spatial truth is enforced externally.
We’re not replacing CAD. We’re using LLMs as idea generators that must be immediately validated by a physics engine. The LLM writes the code, the renderer does the geometry. The model predicts text, but the simulation catches the lies.
This is the unconquered frontier. Not because the models are weak, but because we’ve been asking the wrong question. We wanted a conversational engineer. What we need is a translator that converts words into constraints—and then tests those constraints in reality.
So blow the dust off your AutoCAD. But also hook up your LLM to OpenSCAD. Let the language model suggest, but let the geometry decide. Because the only thing worse than a machine that can’t design is a machine that sounds like it can.
FAQ
Q: Aren't LLMs improving fast? Won't they soon get spatial reasoning?
A: Not by scaling language models. Spatial reasoning requires a world model that understands physics, occlusion, and constraints—none of which emerge from next-token prediction. Even with chain-of-thought, the underlying representation is still text, not geometry. Don't hold your breath.
Q: So what should I do if I need AI help with mechanical design today?
A: Use LLMs for brainstorming, but force them to output executable code (OpenSCAD, build123d, or even Python with collision checks). Then render and simulate the output. The LLM writes the script; the renderer tells you if it's garbage. Never trust a verbal description of a 3D part.
Q: Isn't this just a prompt engineering problem? Couldn't better prompts fix it?
A: No. You can't prompt a model to understand geometry if it has no internal representation of 3D space. Better prompts might produce more plausible nonsense, but they won't magically give the model a physics engine. The only proven workaround is to externalize the reasoning into a computational geometry tool.