Mike Masnick ✅Dec 20, 2022

I asked ChatGPT to write me an entire Seinfeld episode about George getting banned from Twitter. It got the basic beats right... but... wasn't very good at making it funny.

For the final scene I emphasized, please make it funny, and instead of... doing that, it had Elaine tell George he should make sure to be funny on Twitter.

Show threadzalcarikDec 20, 2022

@mmasnick Oof. That's pretty dire.

Can ChatGPT do other sitcom formulas? Curious if this is just all sitcoms being mapped to a common blandness, or if there's something specific about Seinfeld that resists algorithmic decomposition.

Show threadEric William LinDec 20, 2022
@zalcarik @mmasnick I think it’s everything mapped onto a generic statistically plausible median…which means it always has a believable shape, but very mediocre. I think this is generalizable across domains. I experimented a lot with cocktail-related prompts for example…and you basically get a mediocre blandness to everything. It “knows” was a Negroni variations means…but can never really break out and understand how to break patterns.
Show threadEric William Lin
@zalcarik @mmasnick so it’s not just about being “funny”. It’s all domains. Same thing with its sonnets and haikus. Tell it to write a poem and it defaults to trying to rhyme. But often without any understanding of overall rhyme schemes. I wonder if LLMs could be designed to have an understanding of time and chase and effect. (Or have a memory of what it wrote 3 paragraphs earlier)
Dec 20, 2022 at 6:06amMastodon for iOS
Show threademmatonkinDec 20, 2022
@ericwilliamlin @zalcarik @mmasnick
I tried to persuade it to slant rhyme but it was not having it. Villanelles, on the other hand, no worries - as long as you don't look too closely at the result.