> Weirdly it said, "I’ve seen that exact house before on Google Street View when exploring Cairns neighborhoods."
That's slightly creepy!
The anthropomorphisation certainly is weird. But the technical aspect seems even weirder. Did OpenAI really build dedicated tools to have their models train on Google Street View? Or do they have generic technology for browsing complex sites like Street view?
It’s just a hallucination, same idea as o3 claiming that it uses its laptop to mine Bitcoin:
https://transluce.org/investigating-o3-truthfulness
I doubt the model was trained on Street View, but even if it was, LLMs don’t retain any “memory” of how/when they were trained, so any element of truthfulness would be coincidental.
If it's trained on street view data it's not unlikely that the model can associate a particular piece of context to street view. For example, a picture can have telltale signs that street view content has, such as blurred faces and street signs, watermarks, etc.
Even if it's not directly trained on street view data it has probably encountered street view content in it's training dataset.
The training process doesn't preserve information needed for the LLM to infer that. It cannot be anything other than nonsense that sounds plausible, which is what they do best.
I think the test which the OP performed (to pick a random street view and let it pinpoint it) would indicate that it has ingested some kind of information in this regard in a structured manner.
They should definitely add that feature.
Tell it your name and then it just looks you up and street views your house, and puts that all into memory.