Seems like an AI could perform this sort of analysis to geolocate images with far fewer distinct features and no reliance on any prior knowledge, since it would have near omniscient knowledge of all satellite imagery. For example, this seemingly unidentifiable shot of Gwern’s window view [0] contains a nondescript tree next to a body of water. Nigh impossible for a human without any prior knowledge of Gwern’s general location to identify, but potentially trivial for an AI, since it could automatically scour all satellite imagery for trees and shrubbery of similar shape next to a body of water and a house with a three-sash bay window, a task utterly infeasible to do manually.