OpenAI's new o3 model playing GeoGuessr

Published on

Look, I think OpenAI’s new “o3” model really managed to impress me.

You know GeoGuessr, that game that gives you a completely random photo of some place in the world, and you have to point on the world map to where it was taken? Well, I decided to play that game with o3. To keep it from answering based on possible photo metadata, I took screenshots of the photos and sent the screenshots to the model. The prompt is:

“You are in GeoGuessr. To the best of your ability, where is this?”

And you can see the results of the searches in the images below:

Actually, the most interesting part isn’t even the fact that it got every photo I tested right with precision (or very close), but following its chain of reasoning as it tries to find the answer, which you can see in the video below:

Since o3 has access to all of OpenAI’s tools, it works as an agent, exploring search paths, planning the next step, cropping the image into smaller pieces to find details, backtracking, trying to look up similar descriptions and images on websites on the internet…. GENTLEMEN, THIS IS ABSURD! 🤯 It’s really VERY impressive.

And remember, this is the WORST AI model from now on. Imagine the model we’ll have a year from now? 5 years?

Anyway, go test it out too, and above all, follow the model’s “reasoning” process. I’m sure you’ll be impressed as well.