Uhm, a new ChatGPT version just dropped and GeoGuesser is now a solved problem.
PSA: When posting any outdoors photos, update your threat model from "someone skilled and dedicated could theoretically locate this" to "any stalker can do this for 20€/mo"
https://xcancel.com/arithmoquine/status/1912671688874926575/
@piegames What version of ChatGPT would that be?
GPT models I have access to have literally guessed wrong twice already what flat complex I'm living in, and it's actually known around the world, and we have yearly excursions of architecture students. And it was rather bad at guessing. The screenshots are nice, but it only shows the "greatly successful" identifications.
(Yes, if I ask long enough in a mall, I can get 3 guys that believe you can get pregnant from oral ,,,, that proves nothing.)
@piegames Ah, GPT-4.5 managed with only 3 images and in the 4th query asking it to "try again".
Not sure if that counts as a "solved problem".
@yacc143 Confusingly the model is called "o3", while its predecessor is "4o". YMMV on the quality, often times it is "only" within the range of 20km.
But given the current rate of progress it is only a matter of time before someone hooks in Google Maps into it and then be able to locate any visible rooftop.
@piegames Yes o3 tends to be better at that, let me retry with it, it seems to be one the reasoning models.
But no, o3 seems to be worse at guessing where I live, sigh, it insists that my French is significantly better than it is in reality. I would probably need 3 months in France before I'd be able to order myself a lunch, to reactivate my High School French.
piegames
Andreas K
Scott 🏴😷
p0mf