LLMs in some way need to have the ability to learn from the data they see, then weight this appropriately in the model. For the most part we really don't have this. I mean there is the RLHF, but the H is the key that it's human feedback. And even taking this training data and feeding it back in the model is not apt to weight data in such a manner that evolves a common culture over many distinct models.
Now if we see continuous learning models in the future then culture could very well develop.
I tend to agree with how you framed culture, but I was thinking about how culture emerges. Monkey must first climb the stairs to have everyone blasted with water, the outlier act must comes before normalization.
This might be the wrong way of thinking. Culture, mostly, isn't all the things we do different, but the things we harmonize and do the same.
https://www.scientificamerican.com/article/monkey-imitation-...
LLMs in some way need to have the ability to learn from the data they see, then weight this appropriately in the model. For the most part we really don't have this. I mean there is the RLHF, but the H is the key that it's human feedback. And even taking this training data and feeding it back in the model is not apt to weight data in such a manner that evolves a common culture over many distinct models.
Now if we see continuous learning models in the future then culture could very well develop.