It's not difficulty in planning that's responsible for the ubiquitous right angle, it's the tendency of people to like familiar features and want their furnitur

It's not difficulty in planning that's responsible for the ubiquitous right angle, it's the tendency of people to like familiar features and want their furniture to fit (they're also often fond of materials not particularly amenable to 3D printing). Architects would love to design the sort of curvaceous things from concrete, steel and glass they went to architecture school to admire, but the need to sell the houses gets in the way. I'm struggling to see procedural generation of the designs as doing anything but subtract from the ability to conceive 'distinctive style a lot of people like', given that architects are people with likes, can talk to prospective purchasers about their likes and understand at a high level how people might use [or not use] particular design features and what designs regulators might reject. Neural networks, on the other hand, produce outputs which are just a function of their inputs. Procedural building generation is what designers of 3D worlds use for the bits whose appearance they don't care enough about to detail by hand. The one place software design might lead to improved efficiency is randomising features for McMansions on a street from a simple library of squares and automatically generating the docs...