Granted, it wasn't perfect, but it got pretty close, and with some light iterating, the finished result is not bad. Giving it access to prusaslicer as a tool helped it catch some common mistakes like overhang early too.
I'd recommend trying it again, it has gotten significantly better since the last time I tried it about a year ago.
Polygonal geometry is harder for GenAI to do because it’s still largely diffusion based to look like an object and topology is just a polygonal projection over it. It’s roughly equivalent to the diffusion model generating the image inputs to photogrammetry (not quite but a trivialized explanation)
Meanwhile CAD is represented as a series of B-rep surfaces and operations. Both are largely more procedural and more akin to a programming language for models. Therefore it’s easier for an LLM to drive something towards an end result.
Then the CAD engines tessellator generates the polygons, but topology is not super important for CAD.
The best LLM CAD results I’ve seen even in academia are largely simplistic still though. The more data you put in , the better it’ll be able to match that but it’s sort of like giving it a very clear set of functions it can do to make something.
The way it works is the same way all other agentic AI works - feedback loops, and an AI driving a computer. The AI creates a shape in Blender, then takes orthographic view screenshots so it can "see" what it did, double checks its work against what the goal is, fixes errors if there are any, and then repeats until done.
You can find the latest examples on social media or Youtube if you search for keywords Astra, Opus 5.5, and Blender.
This is a notably different approach than many of the existing "text to 3D" or "image to 3D" generators, which train their own machine learning algorithm to output models. Those have also been getting better, but the latest in this field was Astra driving Blender.