The feedback loop they describe sounds a lot like model collapse. They can play whack a mole with the trends they can see, but what about the more subtle forms?
They’re now filtering goblin-related training data, which also tells me that maybe we can use lots of goblin references as a way to opt out of our written content being used to train their models, in our writing and in our code.
totally what I was thinking. what a great way of saying “untold horrors abound”. with so many people treating this shit as an advisor, mentor, psychologist and even partner, i cannot even fathom what weird problems will plague our society without us even realizing.
This is fascinating.