It's my understanding that o3 was trained on multimodal data, including imagery. Is it unreasonable to assume its training data includes images of these exact locations and features? GeoGuesser uses Google Maps, and Google Maps purchases most of its imagery from third-parties these days. If those third parties aren't also selling to all the big AI companies, I'd be very surprised.
> Google Maps purchases most of its imagery from third-parties these days
Maps maybe, but Streetview? Rainbolt just did a video with two Maps PMs recently and it sounds like they still source all their street view themselves considering the special camera and car needed, etc.
Maybe the end-user isn't Google Maps, but TomTom have a pretty comprehensive street-view-ish product for private buyers like car companies, Bing and Apple Maps called MoMa.
I'd be surprised if this building[0] wasn't included in their dataset from every road-side angle possible, alongside every piece of locational metadata imaginable, and I'd be surprised if that dataset hasn't made it into OpenAI's training data - especially when TomTom's relationship to Microsoft, and Microsoft's relationship to OpenAI, is taken into account.
You can upload your own panoramic images to Street View, people do this for hiking trails. But I'm sure 99% of streetview imagery is Google-sourced and Geoguessr might not even use user-submitted imagery.
Trekker coverage is often official, too. I think you are confusing this with photospheres. Also important to note that there is vehicle coverage that is unofficial.
That's true for heavily photographed urban areas. I've tried it on some desert photos (even at identifiable points with human structures) and it basically just guesses a random trailhead in Joshua Tree and makes up a BS explanation for why it matches.
For cities and landscapes in the US it is scary good. I tried about twenty photos from (mostly rural) Central and South America and it is much less effective there w/o a well-known landmark to go off of (likely unsurprisingly). In four cases it got the wrong continent entirely, and in one case guessed as far away as India.
