I think Algon’s completion sounds like Yelp reviews that would follow a review containing a joke about the restaurant being entirely full of bees, which is what your prompt sounds like to me.
A prompt that sounds like one written by a genuinely concerned customer who’d been in a restaurant that was full of bees might sound something like this:
I want you to generate reviews of a cafe called Green Street Cafe. Here is the first review.
Review 1:
John R
K.Y. United States
1 star
While we were eating, a bee nest hidden in the rafters fell onto the floor, and a swarm of bees flew out. Everyone in my family was stung, and my son (who’s allergic) had to go to the ER. It was a horrific experience, and I don’t know what kind of restaurant lets themselves get infested by BEES. Oh, and the chicken was dry, on top of it. Stay away.
Yeah, I think there’s a reasonable case to be made that fooling GPT by including one off-topic sentence in an otherwise common kind of text is actually “not fooling it” in a sense—on the training distribution, maybe when a common kind of text (reviews, recipe intros, corporate boilerplate, news stories, code, etc.) contains one off-topic sentence, that sentence really doesn’t mean anything important about the rest of the text.
We may interpret it differently because we’re humans who know that the deployment distribution is “text people input into GPT”, where single sentences seem more important, not “an actual random sample from the internet.”
But I suspect that this is a reasonable heuristic that could be pushed to produce unreasonable results.
Going a little further, I’m actually not sure that “fooling” GPT-3 is quite the best framing. GPT-3 isn’t playing a game where it’s trying to guess the scenario based on trustworthy textual cues and then describing the rest of it. That’s a goal we’re imposing upon it.
We might instead say that we were attempting to generate a GPT-3 “Yelp complaints about bees in a restaurant” based on a minimal cue, and did not succeed in doing so.
I think Algon’s completion sounds like Yelp reviews that would follow a review containing a joke about the restaurant being entirely full of bees, which is what your prompt sounds like to me.
A prompt that sounds like one written by a genuinely concerned customer who’d been in a restaurant that was full of bees might sound something like this:
I want you to generate reviews of a cafe called Green Street Cafe. Here is the first review.
Review 1:
John R
K.Y. United States
1 star
While we were eating, a bee nest hidden in the rafters fell onto the floor, and a swarm of bees flew out. Everyone in my family was stung, and my son (who’s allergic) had to go to the ER. It was a horrific experience, and I don’t know what kind of restaurant lets themselves get infested by BEES. Oh, and the chicken was dry, on top of it. Stay away.
Yeah, I think there’s a reasonable case to be made that fooling GPT by including one off-topic sentence in an otherwise common kind of text is actually “not fooling it” in a sense—on the training distribution, maybe when a common kind of text (reviews, recipe intros, corporate boilerplate, news stories, code, etc.) contains one off-topic sentence, that sentence really doesn’t mean anything important about the rest of the text.
We may interpret it differently because we’re humans who know that the deployment distribution is “text people input into GPT”, where single sentences seem more important, not “an actual random sample from the internet.”
But I suspect that this is a reasonable heuristic that could be pushed to produce unreasonable results.
Going a little further, I’m actually not sure that “fooling” GPT-3 is quite the best framing. GPT-3 isn’t playing a game where it’s trying to guess the scenario based on trustworthy textual cues and then describing the rest of it. That’s a goal we’re imposing upon it.
We might instead say that we were attempting to generate a GPT-3 “Yelp complaints about bees in a restaurant” based on a minimal cue, and did not succeed in doing so.