Interesting, I wonder if the rugby thing is a common bias. I did find myself in the weights, as the top result. But apparently there are also Australian rugby versions of me!
GP didn't give enough information to know if this was actually a hallucination or not, let alone what type of hallucination. I.e. it's only a hallucination if no rugby player or surgeon is a John Doe, not if John Doe the GP isn't those things.
I wonder how much of hallucinating/"mistakes" in LLMs is because the training data is full of us filling in additional info we humans commonly feel or interpret as implied rather than something which manifests from the architecture of the LLM itself. I assume only a small percentage, but also a non-zero one.
Strange, there’s a neurosurgeon and Australian Rules Football player that share my uncommon name. I already knew about them from googling myself previously. Eerily similar!
Is there any reason to assume it wouldn't be? A lot of training data comes from the open web, after all, and Google also searches Google books, so a Google search is basically a model training data search.
The only interesting thing is how small the models have to be, to lose knowledge of you.