Hacker News Reader

Comment by loupol | Hacker News Reader

I wonder how many images from Street View it has been trained on.

I've seen top Geoguessr players be able to pretty consistently determine a location worldwide after seeing a photo for just one second. So I would assume training an LLM to do the same would definitely be doable.

View on HN

Yep, some CS/AI grads from Stanford trained an AI on loads of Street View images and built a bot that is able to beat some of the best Geoguessr players: https://www.youtube.com/watch?v=ts5lPDV--cU

IIRC it wasn't that impressive in the end as instead of recognizing the places the AI apparently learnt to recognize subtle differences in street view cameras used in different locations? I might be wrong / thinking of the wrong model l and I'm on mobile without my browsing history so hard to check, but I think it was putting a lot of weight on some pixels that are noisy

Top geoguessr players use this technique as well. IIRC rainbolt mentioned that there is a section of a country where the street view camera has a small blemish from a raindrop on the camera so you can instantly tell where you are if you notice that.

From my understanding many of the best players immediately look down to tell what "generation streetview car" they're using, and seem to know what continents/times they're from.

> I would assume training an LLM to do the same would definitely be doable.

I wouldn't be so sure. The reasoning process of Geoguessr pros is symbolic, not statistical inference.

/edit: as other commenters pointed out, something similar was done. While this wasn't an LLM, it was a deep learning model, so not symbolic -> https://www.theregister.com/2023/07/15/pigeon_model_geolocat...

It's been done recently! It's a bit better than (but competitive with) top players.

https://www.youtube.com/watch?v=ts5lPDV--cU

Reply on news.ycombinator.com