I miss that small period of AI where it excelled at making random shit in this vein( see https://x.com/dril_gpt2 and https://www.reddit.com/r/SubSimulatorGPT2/top/?t=all [somewhat nsfw text]).
I'm not really sure why modern AI can't really do stuff like that anymore—I would guess a combination of being whacked with a crowbar to submit to humans (RLHF, but I'm not sure if it affects base models?), alignment stuff, and just being too smart.
Would love to see something like this with GPT-2 and how it compares