• Adalast@lemmy.world
    link
    fedilink
    arrow-up
    4
    ·
    1 year ago

    This comment deserved to be separated from the other discussion. I am studying some LLM stuff as a side project for myself and the author of the book I am reading was discussing the history of AI training a bit in the chapter I was reading. I personally did not realize that LLM models dated back to the very early 200X years. The whole "training on works of art" dates all the way back to the earliest days using non-licensed books and manuscripts in addition to emails, text messages, blog posts, news articles, etc. Scraping whatever content is needed to train an AI from the internet without really worrying about permission is very much so nothing new. It is just something that came to the forefront of the cultural zeitgeist with the release of SD and the clamor of attention it got.

    I think the reason it was never really worried about is precisely how destructive the whole process is. The "Vectorization" step that is common to most if not all AI training algorithms fundamentally disassembles whatever the input is and applies statistical methods to make it something a computer can understand. How many times was each word used, what are the odds of two colors being next to each other, how many times did person A tap their foot? Once this is done, the original work is gone. There are no discernable features of the source material save for perhaps words that are unique to that, but most of the time those are filtered out, so even those are gone. That vector is what the AI is actually trained on, not the original work. All the sources are are chaos to derive statistics from. Nothing more, nothing less.