I mean it is obvious from the get go. Even without the bot training part, what LLMs end up with is some kind of smoothie of the knowledge where all the crunchy

I mean it is obvious from the get go. Even without the bot training part, what LLMs end up with is some kind of smoothie of the knowledge where all the crunchy bits are removed. Like taking the output of a JPEG compression and using it as input, over and over. LLMs representation of our collective knowledge is just going to gradually converge. What I find scary is that this "smoothie" version of our collective knowledge then goes into producing the images and words that surround us, everywhere. Taking away more and more authenticity from us.