Yes! I've also noticed that GPT-3.5 and GPT-4 have somewhat different senses of humor. GPT-3.5 tends to lean more towards absurdism — sometimes it's so bad that it's actually good, but that's not always the case. So, I use both for generating headlines.

However, GPT-3.5 often struggles to generate coherent comment threads, so I use GPT-4 for that, which does the job pretty well.

For images, I simply ask LLM to include a prompt in the img alt tag, and then I call the DALL-E API to generate JPEGs. But honestly, DALL-E isn't great; it would be much better to use StableDiffusion or Midjourney, but I'm too lazy to integrate them (does Midjourney even have a public API?)

If you want to see the prompts, here is the code: https://github.com/crackernews/crackernews.github.io