I've been thinking a lot lately about the recent wave of AI tools promising to turn text descriptions into finished CAD models. Companies like Leo AI are banking on the idea that you can describe a part in plain English and have a model spit out a ready-to-manufacture design. To me, this feels like we're missing something fundamental about why CAD exists in the first place.
The History CAD Is Trying to Tell You
Before CAD was CAD, we had drawings. Professional engineers would spend hours creating detailed blueprints by hand. These weren't just pretty pictures; they were the only way to communicate complex geometric and dimensional information to manufacturers. Here's the key insight: language alone wasn't enough. An engineer couldn't write a paragraph describing a part and trust a manufacturer to build it correctly. You needed visual information paired with precise mathematical data.
The fact that we moved from drawings to digital CAD to full 3D models tells you something important: we've always been moving toward richer, more complete ways of capturing design intent.
The Feel of a Curve
I've talked to industrial designers and CAD drafters who spend hours getting a single curve just right. Sometimes it's for aesthetic purposes, sometimes for ergonomics. You can't really capture that in a sentence; it's more about the feel and the vibe of something. That's the kind of complexity we're talking about.
Precision Is Hard Even for 2D
This brings me to something I noticed while spending time with generative image tools. When you care about precision and consistency, these tools struggle. I've found myself spending hours trying to coax Nanobanana or Sora into producing exactly what I need; the attention to detail required often feels like I'm working against the tool rather than with it. And that's with 2D images. CAD demands something exponentially harder: mathematical precision combined with aesthetic intention, all captured in dimensional space.
CAD Is a Compressed Language
Here's another way to think about it: math is a language, but it's a compressed one. Try describing the Navier-Stokes equation in English; you'll need paragraphs to capture what a few symbols convey perfectly. A CAD model does the same thing. It's a domain-specific language optimized for spatial and manufacturing complexity in ways natural language simply cannot be.
Where LLMs Actually Fit
So where do LLMs actually fit? I think the answer is in knowledge and decision support, not design automation. If you're a junior engineer working on high-speed PCIe Express routing, an LLM can help you understand differential pair routing, impedance control, and EMI considerations. It can guide you through regulatory checklists, help you think about tolerancing, and accelerate your learning so you can make better decisions in CAD. That's where language excels.
Generative CAD Isn't Impossible — LLMs Just Aren't the Tool
But here's where my thinking gets more speculative: I don't think generative CAD is impossible. I just don't think LLMs are the right tool for it. Diffusion models transformed image generation; transformers revolutionized language. There's probably some fundamentally different technology needed to properly generate CAD models with the precision and control that manufacturing demands. I'm not an AI researcher, so I can't say what that technology looks like. But I think it's worth asking: what would it actually take to build generative CAD that works? And does that technology exist yet?
For hardware engineers and teams, I think the practical question is simpler: how do we use LLMs to actually help us right now? And then separately, what underlying technology would be required to solve the generative CAD problem properly? Those feel like two questions worth keeping in mind.