Turns out floorplans is the perfect case study for why AI doesn't work, on several levels even.
-
Turns out floorplans is the perfect case study for why AI doesn't work, on several levels even.
Like, why too many bathrooms? Because in a floorplan, the bathroom is relatively information-dense, so statistically there's an overweight of bathroom data in the training set.
Also, a plan is a concept by definition & you can't build that from the smallest unit up; planning is literally the opposite of that.
Et cetera.
-
J jwcph@helvede.net shared this topic
-
Turns out floorplans is the perfect case study for why AI doesn't work, on several levels even.
Like, why too many bathrooms? Because in a floorplan, the bathroom is relatively information-dense, so statistically there's an overweight of bathroom data in the training set.
Also, a plan is a concept by definition & you can't build that from the smallest unit up; planning is literally the opposite of that.
Et cetera.
@jwcph even if the plan would work in reality, can you imagine how an A.I. project manager would f*up the logistics of building and construction?
That's a complex thing to do, which needs reasonable people constantly doing damage control. -
Turns out floorplans is the perfect case study for why AI doesn't work, on several levels even.
Like, why too many bathrooms? Because in a floorplan, the bathroom is relatively information-dense, so statistically there's an overweight of bathroom data in the training set.
Also, a plan is a concept by definition & you can't build that from the smallest unit up; planning is literally the opposite of that.
Et cetera.
@jwcph
Having trained as an architect, these floor plans are
. You really need to think four-dimensionally, so not just in volume but also in sequence of how people use space. And you need to be able to „think around the corner“ (literally). How things are joined matters. These plans don’t even work as a draft that could be made to work with some human oversight. They lack the essence.
Doesn’t mean, that it won’t „learn“ in the upcoming years. But it’s certainly not anywhere close today. -
@jwcph
Having trained as an architect, these floor plans are. You really need to think four-dimensionally, so not just in volume but also in sequence of how people use space. And you need to be able to „think around the corner“ (literally). How things are joined matters. These plans don’t even work as a draft that could be made to work with some human oversight. They lack the essence.
Doesn’t mean, that it won’t „learn“ in the upcoming years. But it’s certainly not anywhere close today.@samy Yes, it does mean it won't learn - because as Nicole says, you can't get to the overall concept of a floor plan, including things like the 4th dimension you mention, by applying statistics at a pixel-by-pixel level. It just isn't possible.
-
Turns out floorplans is the perfect case study for why AI doesn't work, on several levels even.
Like, why too many bathrooms? Because in a floorplan, the bathroom is relatively information-dense, so statistically there's an overweight of bathroom data in the training set.
Also, a plan is a concept by definition & you can't build that from the smallest unit up; planning is literally the opposite of that.
Et cetera.
@jwcph would be interesting to know which houses the model is trained on - badly designed McMansion disasters with an over abundance of bathrooms, kitchens, gapping three storey entrances, and roof top swimming pools will nicely skew the statistics
-
@jwcph would be interesting to know which houses the model is trained on - badly designed McMansion disasters with an over abundance of bathrooms, kitchens, gapping three storey entrances, and roof top swimming pools will nicely skew the statistics
@iain I'm willing to bet this is just regular GenAI slop - they weren't trained on houses at all, just on pictures in general.
