Design Intent & Vibe-CAD
I got started with 3D modeling in maybe 2018 or 2019. With TinkerCAD. Didn’t even know the word parametric back then. When I wanted to make something, I would start with primitive shapes like a rectangle or circle and cut/extrude them until I got I wanted. To be fair, I don’t think TinkerCAD offered anything beyond that anyway.
In my first year of University, I decided to move away from TinkerCAD and learn Fusion360 instead, since it was closer to a proper CAD tool. Since I preferred learning by tinkering and only look up tutorials when I needed to learn something specific, I didn’t really understand the importance of parametrization or design intent at that time. My workflow was very messy, a lot of random extrude/cut operations here and there. Basically the virtual equivalent of taking a hammer and beating a metal block with it randomly until it looked right.
It worked for me - until I would have to change something in the design. Due to the lack of well defined relationships and constraints, changing anything in an earlier stage was very hard.
My CAD game got better in the last two years. Finally learned about the importance of a clean workflow and making reusable, modifiable CADs. Still learning but yeah, it’s not as bad as it used to be.
Sorry about the long preamble. I just wanted to say that there is a difference between making a CAD model (for engineering purposes) with vibes (me in 2022) and making it in a thoughtful way with a properly laid out execution plan (hopefully me now). The end result can be the same for both. But the latter is what engineers need. Or, that’s what I think they need.
AI has been overhyped for a long time. Every time I see some AI tool claiming they will come for the job of mechanical engineers, I get curious because I don’t want my job to be replaced before I even get one. However, a lot of the tools don’t live up to the promise. Text-to-CAD tools where you say “Jarvis, build me an engine” and it spits out a 3D model of an engine may look impressive in demos. And possibly impress a lot of people. But it doesn’t really matter if the design intent element is missing in the workflow. For design element, there needs to be context. A human knows the context because they know what they are designing for, and possibly have (assuming a professional environment) clearly documented specifications that the design has to maintain.
It’s not hard to make those specifications AI-digestible and feed that into the context window, from my understanding. But apart from specifications of the current project, human engineers also have access to more context from their experience. In order for an AI agent to even have a chance of replicating the thought process of a human, we need to document that experience and feed it as well, and train it on that data. Sure, someone can make an AI agent and make it train on general experience data. But if it’s not your company’s data it was trained on, it’s not really that much different than going online and hiring a freelance worker who have zero experience with your company but has a lot of general world.
I think we have seen a lot of development with AI models in the last one year. And at the moment, plenty of them are capable enough for doing serious work. But I would like to see the next updates focus more on creating AI-native specifications, because even a smart person (comparing the AI agent to a human here) with no context is worse than a comparatively incompetent person with more context.
Update 03.07.2025 - Looks like Fable 5 is back again. It is still absurdly expensive though. I wonder whether cheaper models with better feedback loop can outperform Fable in CAD.