o3 is scarily good at geo-guessing but can't work out what's what on a pixelated screen. That doesn't make sense to me. Maybe a lack of pixelated training data?