I even typed in: "person who sings EP MyWorld" and got the same answer. Given how much data there is about him (the infamous Twitter scaling problems), I don't think it's necessarily a good test. There's gonna be tons of relations in the corpus for him.
For me the remarkable part is that it’s emergent behavior in large scale language models.
The default sampling parameters are really not that hot for cases where there is nearly a unique correct answer-- bad luck will force to to ask wrong.
From that prompt I (a human) don't know if I should expect something like "Albert Einstein" or the straight answer, so higher temperature sampling can risk kicking it down a silly path.
A: Justin Bieber.
So that suggests that the prompt and temperature (etc.) are relevant to getting the kind of behavior you want. But it didn't seem to have any trouble getting this right the first time with the appropriate preset.
> A: Justin Bieber is a singer, songwriter, and actor discovered as a teenager by talent manager Scooter Braun.
or
> Q: What famous singer was discovered as a teenager by talent manager Scooter Braun?
> A: Justin Bieber was discovered as a teenager by talent manager Scooter Braun.
It's really good when in the factual question answering mode.
> Q: What famous singer was discovered as a teenager by talent manager Scooter Braun?
> A: Justin Bieber was discovered as a teenager by talent manager Scooter Braun.
> Q: What country is he from?
> A: He is from Canada.
> Q: Who's the prime minister of that country?
> A: Justin Trudeau is the prime minister of Canada.
> Q: Who was his father?
> A: Pierre Trudeau was his father.
> Q: Why was Pierre Trudeau famous?
> A: Pierre Trudeau was famous because he was the prime minister of Canada.
> Q: When did he take office?
> A: He took office in 1968.
> Q: What was the year before that?
> A: The year before that was 1967.
> Q: What was the summer of that year known as?
> A: The summer of that year was known as the "Summer of Love".
Given we really don't know how it's all working, it's safe to say that prompt composition is an artform that one can develop a skill at.
I was on a discord server with some other folks and one of them had access to the GPT-3 api. We provided a prompt similar to the following:
At first pass it gave some weird answer. So we deleted the answer and tried again. The second time it wrote 'Justin Drew Bieber'.That chunk of biographical info is from his Wikipedia page, but it's really kind of remarkable to me that it was able to really even understand the question, let alone mine its relationships and cough up the right answer.