Is it good, or are those photos already in the training data that the model was trained on?

Would be interesting to see how well it would fare on five photos from different years, using pictures that have never previously been uploaded anywhere.

I would bet it would still do well. You can give a frontier LLM photos that it has never seen before and it will reason about them incredibly accurately, since is is trained on, well, a lot of other similar photos. You can easily try it yourself, give o3 a photo you have taken yourself of a random place, it will crop, zoom and analyze different sections of the photo until it gets most of the time the exact coordinates of where the photo was taken.

> it will crop, zoom and analyze different sections of the photo until it gets most of the time the exact coordinates of where the photo was taken.

I will give it a try myself too. But all my photos that I take with my iPhone have date and GPS location embedded in the metadata so I will first have to transfer the pictures to my computer and strip them of metadata.

You can just take a screenshot of them. Not the cleanest way to do it but probably the fastest if you don't already have special tools on your phone.

> are those photos already in the training data that the model was trained on?

The photos are of course in the training data. Most humans too, would have seen a large fraction of these photos before.

But associating it with a year is where the magic happens.

You are dismissing a technology capability that was science fiction a year or two ago.