Royalties for inference are unrealistic in a way that even royalties for training aren't.
The LLaMA models were released openly. Copies exist everywhere in the world. You aren't going to be able to charge someone for running `llama.cpp`; a court order ceases to have practical relevance at that point.
Inference might be unreasonable for a royalty agreement, but, in assessing damages, it is certainly relevant.
"I made enough copies for everyone" isn't a valid defense for copyright infringement.
These models can provide citations so I don't see why they can't tick a royalty owed. I'm sure many here could help build this pipeline.
First, LLMs do not reliably cite works. They are not looking things up in a database and repeating them. I think this false idea occurs a lot in people who don't understand what LLMs are or how they work.
Second, royalties are not required to cite a source.
Can you imagine how disastrous it would be to everything from news reporting to scientific publishing if that was the case?
Yeah well then I want my robot running this crap locally in its brain so I can get it to farm my two acres and haul water for me and I'll unplug from the rest of this nonsense going forward lol.
... LLMs cannot reliably provide citations. If you ask for citations, and the model did not use a web search tool, then whatever "citations" you receive are unreliable. Please do not trust these models to be honest. Just because they can discuss a topic doesn't mean they "know" where the knowledge came from in the same way that you don't need to have studied physics to catch a ball.