One thing that screws it up is ironically the memory across conversations. I gave it some _really_ obscure photos from some godforsaken ass cracks of the world I've been to, and it was able to guess most of them correctly. However, in its reasoning trace I saw that it's having a heck of a time letting go of the idea that the photo N is not roughly from the same location as the previous photos. What's even more impressive, even when it guesses incorrectly, it can often follow up and guess correctly with minimal hints. And it reasons about it much like a human would, and searches for the same things a human would. Note: I used o3, YMMV with a smaller/weaker model.