Ask HN: Is humor the next frontier for LLMs?

1 point

a year ago

Not so long ago, we used to think that machines could be considered intelligent/creative/conscious/<some attribute deemed quintessentially human> if they could write a sonnet/summarize a book/paint a portrait/tell a joke, or whatever. Back then, all these tasks seemed technologically far beyond our reach.

Today, some of those have been achieved: even if they don't produce truly _good_ poems, LLMs are already better than the median human effort in poetry [1]. The same goes for paintings, essays, and so on.

Yet humor remains a tough nut to crack. LLM-generated jokes often suffer from:

  - off rhythms or pacing
  - missing or telegraphed punchlines
  - unintentionally absurd twists
  - over-explanation of their own process
  - reliance on weak wordplay

They do a decent job at generating satirical, Onion-style headlines, but longer humorous narratives still elude them.

So, could humor be the next frontier for LLMs? That is, not as evidence that they are intelligent/creative/conscious/<some attribute deemed quintessentially human>, but as the next area that is easy for humans yet hard for machines. Or am I just a bad prompter?

[1] Humans actually seem to prefer AI-generated poetry: https://www.nature.com/articles/s41598-024-76900-1