We used to use markov chain generators of various types to spit out absolute insanity, in the early years of the century, and this is all I ever wanted from LLMs. I bet if you used a slightly older model for your text generation, you'd get really clean output. Sonnet 4.whatever, I think, actually does a superb job at text generation, despite it being only an ok coding model.
Extremely small or very heavily quantized LLMs can generate hilarious results. You can also get wild shit if you dial temperature really high.
Author here. Surprised to see this posted on HN. we had a Markov chain irc bot who was quite a jokester
https://charlesleifer.com/blog/building-markov-chain-irc-bot...