@bonifartius

> llm produces semi random combinations

yeah. this is something a lot of people have a hard time wrapping their head around. it's just a fancy autocomplete, which in the most basic sense can be implemented as a spreadsheet with a minimal algorithm as the front-end. nothing "intelligent" about it. but that also doesn't negate it's usefulness. however quasi-deterministic it may be, it's still a startlingly convincing simulacrum. one thing that I think of often in this context is the story of the golem of prague, except the modern incarnation isn't likely to come with an off switch (מֵת).

> if they were trained locally.

would be nice. I've created toy models this way purely for educational purposes, but to do anything particularly useful it seems like I'd need at least a couple $100K or an order of magnitude more, and dedicated access to a data centre and team of other engineers. so, I rent instead, and use RAG with strong guardrails for domain specific knowledge. it mostly does the job I want it to, albeit with a lot of back and forth and hand holding to avoid it producing a pile of incoherent slop.