People tried VLMs on "closed set" GeoGuessr-type tasks - i.e. non-Street View photos in similar style, not published anywhere.
They still kicked ass.
It seems like those AIs just have an awful lot of location familiarity. They've seen enough tagged photos to be able to pick up on the patterns, and generalize that to kicking ass at GeoGuessr.
Comments
People tried VLMs on "closed set" GeoGuessr-type tasks - i.e. non-Street View photos in similar style, not published anywhere.
They still kicked ass.
It seems like those AIs just have an awful lot of location familiarity. They've seen enough tagged photos to be able to pick up on the patterns, and generalize that to kicking ass at GeoGuessr.