TL;DR
"LLMs are like music samplers instead of mp3 pirates, publishers are doomed from things other than LLMs, creativity is the loser with LLMs, it's all about the relationships between words."
First, LLMs and book are not at all like downloading and music. A better comparison would be sampling music to create new music. Imagine where the music publishers would be today if 20(?) years ago people could've lifted instrument/vocal tracks from a bunch of songs, mix them into new songs, and sell the new songs without giving any money back to the original artists? Just imagine a seamless duo of Alice Cooper & Karen Carpenter singing an Easy Listening version of "Bulls on Parade"
The copyright battles between the samplers and the music publishers were easier than the coming copyright battles between LLM producers and book publishers, the written word space is so much larger than the music space.
I'm not actually worried about the publishers, with ebooks and self-publishing getting easier their current marketing model has a limited lifespan. The danger of LLMs is their lack of creativity. They can 'fake' creativity if they have a large enough data set to come up with a response that looks creative, hence the need to consume huge amounts of creatively generated content. This is also why LLMs consuming LLM generated content is bad for the LLM, it has 0 creative content.
The judge in the copyright case of LLM vs Publishers/authors missed the point. LLMs, and their non-text base cousins, consume and store structured data, like text, in such a way as to make the structured data easier to scan and create relationships to other consumed structured data. In effect they are storing the entire, or at least most of, the content of a consumed book in a database, probably in some form of a relationship graph, and then using that data to answer questions from the public for free or for a cost.
This could be considered to be like a model of a research librarian except that the LLM is answering multiple questions at the same time. If multiple questions use the same relationship edge then this is like having multiple librarians using the same copy of the book that created that relationship. This is not yet a problem as the librarians are very fast and know exactly where in the book they need to look and can put the book back very quickly. However since a single book will generate a multitude of relationships and it's likely that for a given question if one relationship from a book is used, more than one will be used. With a large contextual memory space available, generally with more dollar cost, more of those relationships can be tracked. This leads to the idea that effectively each question that uses a relationship essentially creates a copy of the relationship in essence creating a copy of the book that generated the relationship. Since with a large consumption of texts about a board range of subjects will increase the chance that more than one text with create the same relationship. But if the LLM doesn't track how many texts contributed to the relationship that acts as a multiplier to increase the number of virtual texts are being created.
This exposes the actual value a text, or piece of structured data has to an LLM. It is not the words of the text that are important, it is the relationship between those words that is important. Want to poison an LLM, feed it a large text composed of grammatically correct sentences of random words.
