i still havent seen an llm invent new tokens/words over the course of a session? the novelty remains a structural problem with how they work and what the weigh

i still havent seen an llm invent new tokens/words over the course of a session? the novelty remains a structural problem with how they work and what the weights try to predict