The AI GeoGuessr
Geospot Infinity is a fascinating experiment in OSINT or AI geolocation. The web app lets you upload a photograph, and in seconds it suggests ten possible locations where the picture might have been taken. The most likely match appears at the top of the list, while the remaining nine give alternative guesses.
Built by Surya Dantuluri, Geospot Infinity combines a vision-language model (GeoCLIP) with a small neural network that re-ranks those ten guesses to try to find the best match. In theory, the model can even learn from user interactions - improving its accuracy every time someone picks the correct spot.
However, early testing revealed an interesting quirk: most users simply click the first suggestion, whether it’s right or wrong. This makes it hard for the system to learn accurate feedback, and sometimes results in amusingly misplaced guesses.
To address this, Dantuluri replaced the original reinforcement learning setup with a Direct Preference Optimization (DPO) approach. Instead of rewarding every click equally, DPO compares the clicked location with the unclicked ones and trains the model to prefer whichever the user chose. This makes the feedback signal clearer — especially when users scroll through multiple options before selecting one. In those cases, the model learns that the chosen location was better than all the ones above it, giving it a stronger and more meaningful update. The result is a system that can gradually improve its rankings based on genuine user preferences rather than casual first clicks.
From my limited testing, Geospot Infinity produced mixed results. My first photograph - of Leadenhall Market, probably one of London’s most photographed locations - was repeatedly identified as being inside the Natural History Museum on the other side of the city. I then tried a picture of Tower Bridge, and this time the correct location was identified immediately. Finally, I tested a less touristy location with a photo taken in Epsom. Geospot Infinity first placed it somewhere in Kent, then in Sussex, and finally in Reigate, around ten miles away.
Also See
Bellingcat's Geolocation Tools
Comments