> At my first job out of grad school, I tried to convince my manager that GPT-3 could be used to figure out whether or not a building had air-conditioning vents.

I cannot tell whether that's a joke, but I'm very interested if it's serious

My colleagues definitely thought it was a joke at the time.

We had this project (all public research) to classify buildings and identify their different subsystems (e.g. load-bearing structure, roof type, ventilation type) to figure out the expected casualties if there was a WMD event of some type. We could get decent data for much of the world, but for some places we had literally nothing beyond a tiny picture of it from satellite imagery.

I had been playing with using GPT-3 to try to have it autocomplete forms like the following. This was 2021 before we had good APIs for instruct models, so this was just straight up letting the LLM regurgitate after pretraining. Here was the type of prompt we used:

""" Engineering building report for building located at 123, X Street, Knoxville TN Prepared by Benjamin Lee, FE --- Building footprint area: 1200 m2 Roof type: built-up roofing Facade material: brick HVAC present: """

Surprisingly (at the time), this was a decent prior. You could also add all sorts of one-off points of interest and amenities like swimming pools and other trivia to help guide the conditional probabilities.