Rather than feed his AI thousands of elementary school floorplans and then ask it to produce a new one that fits in with the others, Simon gave the computer one floorplan and asked it to keep the same rooms in the same sizes, but rearrange the layout, first to minimize the amount of building materials and the time it takes to get between classrooms, and then to minimize how long it takes anyone to reach a fire exit.
|Fig 1 - The neural-network's two final designs from Joel Simon|
The structures the computer came up with are probably impossible to build, and would likely feel deeply unsettling to be in. He says "I have very mixed feelings about this project. ... By not obeying any laws of architecture or design, it made the results very hard to evaluate." He could probably get a viable building if he also demanded that the neural-net only draw rectangular rooms and uniform hallway widths. Instead, he ended up with a lot of teardrop, hexagonal, and diamond-shaped rooms;** hallways that get narrower the deeper into the building you go (because the expected foot traffic diminishes as people enter shallower rooms); and hardly any windows. He explains: "Windows were also experimented with ... this led to many interior courtyards."+
Joel Simon's elementary schools don't look like any buildings that actually exist, but they do look like they'd make pretty good dungeons. The numerous branching paths and dead-ends would create a non-linear exploration experience, and the frequent use of room-to-room doorways actually creates a number of loops despite the hallways never reconnecting.
I imagine that the green classrooms would be standard dungeon rooms, while the mustard-colored teachers' lounges are caches hidden by secret doors, the red admin rooms and light blue facilities seem like lairs or faction strongholds, while the lavenders spaces are "specials", and the playground is a garden. The organic, space-saving and material-reducing logic of the place even makes a kind of sense if we imagine that it's underground. (Other explanations could be a building constructed by fungal / alien minds, a building rearranged by evil magic, a memory palace as it exists within the subconscious of the victim of your heist, or the Dark Hyrule / Upside Down counterpart to an ordinary building.)
|Fig 2 - The neural-network's design process from Joel Simon|
* I feel like there is some collective ritual of reassuring ourselves that our jobs are still safe among members of the precariat and various white-collar professions, where we read these things with bated breath and then have a good long post-panic laugh after seeing that neural-nets are still so bad at designing anything. I also feel like that will probably change within my lifetime. I suspect that someday we'll all look back fondly at a time when a computer thought "grass bat" and "turdly" were good names for a paint colors. I can't tell you how relieved I feel whenever the all-seeing eye of surveillance-driven internet advertising offers me something ludicrously inappropriate. It's going to be disquieting when internet ads inevitably learn how to offer me things I might genuinely want.
** I expected circular rooms, and you can see that they are circular in the preliminary floorplan, but then the rooms expanded to fill all the unused space between the circles, leading to a number of odd, hard-to-describe room shapes.
+ I would have liked to see one of the floorplans with an interior courtyard. Again, it's something that you would never build in a real school, but that fits the the underground dungeon aesthetic. Then again, as a child, I used to want to live in a Victorian house that took up the outer ring of a 3x3 grid, with the center square being an interior arboretum - so who knows, maybe AI architects will give us the courtyards we didn't know we wanted.