> Claude just produced a 5-page report for me that uses the adverb "therefore" exactly 16 times.

I might be imagining it, but I really feel that the quality of the writing coming out of LLMs is deteriorating. You went from models like ChatGPT 3.5 that sometimes produced literal gibberish to things like ChatGPT 4.0o that were generating stuff that was both coherent and stylistically consistent and now to the current generation of models that don't normally produce large errors but often sound a bit weird.

I wonder is this is down to the models simply getting huge or that there is now so much LLM generated stuff on the net that they are trying to train on their own output.