Who are three famous people named Zora?
On the other hand, you could deny that "famous" has any specific standard, and there are a lot of people on this page:
https://en.wikipedia.org/wiki/Zora_(name)
It would be extremely easy to avoid agreeing with my personal definition of fame, for a human.
It replied:
Zora Neale Hurston: Zora Neale Hurston was an American folklorist, anthropologist, and author who was a key figure in the Harlem Renaissance. She is best known for her novels "Their Eyes Were Watching God" and "Moses, Man of the Mountain," as well as her work as a folklorist, collecting and preserving African American folktales and oral histories.
Zora Arkus-Duntov: Zora Arkus-Duntov was a Belgian-born engineer who is known as the "father of the Chevrolet Corvette." He played a key role in the development of the iconic sports car and is credited with introducing many of the design features that made it successful.
Zora LaVey: Zora LaVey was an American burlesque dancer, model, and actress. She was known for her exotic dance routines and her appearances in films and television shows in the 1950s and 1960s.
So I wrote: Is Zora LaVey a fictional character?
And it replied by affirming they are a real person, and regurgitating the paragraph above, with some slight embellishments.
But starting a new session and asking the same question produces: "It is not clear who or what you are referring to with the name "Zora LaVey." There is no widely known public figure or fictional character by this name. It is possible that you may be mistaken or that this person is relatively unknown or obscure."
Among other things, it seems to me significant that while it seems "good" at BS'ing, it didn't BS by choosing an actual person who was on the Wikipedia page and explicitly or implicitly disagreeing about who is famous. There is some general principle of how it works that captures this behavior, I feel certain. Obvious if it were stated clearly.
1 comment
[ 3.9 ms ] story [ 13.1 ms ] threadIt is statistical completion, but working at the nth order. What seems most likely to come next? An enormous number of possible responses are generated based on the prompt, and then tested for how likely it seems like everything else in its training data. A possible completion that is a fake biography of a person but which looks very much like a real biography, scores quite highly on the metric. ChatGPT has also been trained by feedback, with fake but convincing syntheses being penalised, and responses saying it doesn't know or using accurate facts being preferred. And so it also does the same testing process with the "how likely is it my trainer will like this completion?" metric. That explains the behaviour in the second session. In the first, it slipped by that metric.