Not necessarily, the LLMs used today are far from just simple models of written information on the internet. They use in-house data they wrote themselves, and R

Not necessarily, the LLMs used today are far from just simple models of written information on the internet. They use in-house data they wrote themselves, and RLHF/DPO where it's effectively training on its own data to optimize for human preference. If sampling with high enough temperature for this, it could theoretically bring out entirely new unseen forms of speech as long as people express their preference for it via the user interface