My experience is it's short to get a working tool and long to finish. Many 3D modulers (blender for example) have literally 1000s of features. Yes, you might not need all of them but you probably need 100s, each with quirks and UI/UX decisions. Just iterating on camera controls can take days. You get an orbit camera working via a prompt. But then find it only zooms in/out so far because the LLM decided that was ideal. So you prompt to fix that. Then you find moving the camera is not zoom aware so it moves in meters per frame and when you're zoomed in things move too faster. So you prompt to fix that. Repeat 10000x. Every feature, extrude, bevel, chamfer, subdivide, select vertices, select faces, select edges, etc etc all require this handholding and finese.
At least that's my experience.
Small tools with few features is great. a 3D modeler is not that.
LLMs also seem prone to making weird decisions that no human would have done. You can very quickly have a broad implementation - often with features that are great and that you didn't even know you wanted - but when you dig into the features, the weirdness reveals itself bit by bit. If you were developing by hand you would have used a sensible range for zoom from the start, but the LLM doesn't seem to work that way.
I suspect people saying "everyone will vibe code all their software" have never actually tried to vibe code a meaningfully complex application.
Yep having the same experience right now with some curve/spline tessellation into triangle strips. There are many degenerate edge cases and when you ask the "ai" to handle one it breaks something else etc. Makes me think that if I had just hand coded it myself I would intimately know the code and get it working faster than prompting. Gah.
> Yep having the same experience right now with some curve/spline tessellation into triangle strips. There are many degenerate edge cases and when you ask the "ai" to handle one it breaks something else etc. Makes me think that if I had just hand coded it myself I would intimately know the code and get it working faster than prompting. Gah.
It gets so much worse. At some point you start hitting edge cases where there is no clear right answer and you have to come up with a kernel wide coherent approach to things like 1ulp differences. Parasolid is full of decades of kludges on top of kludges and giant switch statements trying to preserve backwards compatibility as they explored the space.