> But if you could send an LLM back to 1950 and train it on the pre-existing literature, you couldn't have it generate The Shining, A Scanner Darkly, Piranesi, Perdido Street Station, The April Witch, or Starship Troopers.

Why? I don't see a clear reason why it couldn't that does not invoke the mystical "human spirit".

To be clear, I don't expect a single LLM to create such a book, at least not in anything resembling current models, but I can absolutely imagine a long multi-agent setup that would. Given enough money, I'd set up a mechanism whereby author AIs generate books, separate agents act as editors and critics to iterate on a small fraction of the best of these drafts to polish them, and then publisher AIs choose an even more minute fraction of these to publish into "the zeitgeist", which will then have other agents offer further commentary on these books, which would then feed into the context (or even fine-tuning/training) of the next generation of LLMs. I expect that a sequence of e.g. ten generations of swarms of AIs, each perhaps of ten thousand agents, could arrive at a ground-breaking book with actual literary merit.

Give me the budget to prove me wrong :)

This is similar to code factories.

In my company I have a very good conversational dynamic with my team. So in my first forays into LLM usage (the past couple years), I began trying to designate models as team members and mimic similar conversations and decision making processes that we have at work, just as an experiment. Somebody to brainstorm together who likes going on tangents, somebody who unapologetically removes unnecessary cruft and cuts scope, a few other people with specific tendencies and personalities.

The results were completely awful. LLMs couldn't even remotely mimic these personalities. The conversation was all trying to please no matter how contentious I tried to instruct them to be. LLMs just seem to struggle a lot with arguing and being contentious. Which probably makes sense based on training.

I feel like these types of factories in practice don't work out too well.

Yet.