I don’t think it’s fair to say LLMs are “statistically plausible” text generators. It ignores the emergent world model that enables them to generalise outside o

I don’t think it’s fair to say LLMs are “statistically plausible” text generators. It ignores the emergent world model that enables them to generalise outside of their training set. LLMs aren’t a lookup table, and they’re not n-grams.