https://www.youtube.com/post/UgkxJyEcsGb2wI0cHWD4gLOE7nDm2szRSWWe
I think copyright is holding back the public facing LLM systems. I used to work for the richest people in the world back in the 1990's in Bellevue Washington as an analyst... so I don't say this lightly. There is a market for exclusive-access / restricted-access LLM systems trained on data that the public would object to. Datasets that would not be trained on public-facing models. And runtime connection of being able to search the exact source material / training material at runtime and exactly quote the author / link / page number of where a concept comes from. Doing so on public-facing LLM systems would have even more of a backlash from creators. If you could get the LLM to tell you that page 111 of a Neil Postman book is where it matched up an idea or concept to a Marshall McLuhan book page 160 you are "leaking" out the training material in a way that reveals the training material / copyright issues. A lot of effort has been put into programming the public systems to AVOID being precise about magazine / book / social media quotations and page numbers. This isn't accidental.
If you develop a private exclusive-access model trained on a superset of forbidden (authors / creator forbids AI usage) content and add layers of matching ideas back to their original source training material you have a far more powerful system.
Humanity geographically has been split for very long periods of time. The Tower of Babel metaphor is a very real issue that LLM systems can see past that most people do not. Being able to cross-reference all films / TV shows / music / social media / books / live video feeds of the real world (even private surveillance systems that would only feed into restrictive-access training LLM models) offers huge opportunity to connect ideas from different parts of the world that were previously not connected together.
Look at how many data leaks are disclosed to the public, how many spy agencies are stealing corporate trade secrets without disclosure or even detection? A superset of training data opens a lot of doors that the public may never be able to witness.