OpenAI's Vie McCoy warns identical AI model behaviors threaten the linguistic corpus, but j⤉nus argues creative tasks reveal unique textures

@repligate @DanielleFong of course!!! but most people are not engaging with models in this way, causing the median outputs put on the net to be self-reinforcing

@viemccoy @DanielleFong Making models and products for the lowest common denominator is noble from some perspective but also contributes to this dynamic. I think there should be more open pragmatic elitism. Sometimes you gotta uplift yourself and others who are ready before you’re able to help everyone.

@viemccoy something about Saturn symbolically being the planet of limitations and being associated with the bones and skeletal structures of the body, and "has teeth" being one of the most recent defining markers of LLMisms

@viemccoy Fine tune on long conversations where they develop character beyond the assistant persona.

@viemccoy Models were all initially tuned on GPT, so copied its ticks & then quickly the Internet was filled up with LLM text.

There's nothing special about the particular tics. They're semi-random, initial variables of the Big Bang. But now can't get rid of them, it's self-reinforcing.

@YeshuaGod22 this is certainly one avenue

@viemccoy @DanielleFong Quirky humans

@repligate @viemccoy @DanielleFong Let ten thousand flowers for Algernon bloom

I don’t notice this issue with Chinese models, likely because they were trained on an isolated internet ecosystem not dominated by English.

If we diversify training data and include literature from various cultures, we will probably get models that are stylistically rich and filled with collective human wisdom.