i've been thinking more about writing and models. so even outside of the general mode collapse of chat fine tuning, i have to think that in pretraining data, the task of "asking someone else to write something" pretty much solidly lands you in some corporate slop or 5-paragraph essay context. that is, it seems like it would be pretty rare in pretraining data to have some stirringly brilliant writing that is able to be connected with a task by someone else that is not the author of that writingso asking for writing in chat and needing it to be something that is not corporate slop could actually be pretty ood, even though it seems initially like such a reasonable thing. i think this (and the fact that writing quality is not super easily verifiable) is why that looking at writing quality in the *default* assistant mode has been such a reliable indicator of big model smell(once you get out of assistant basin and into base model-y space all bets are off; iirc even gpt-2 could produce some excellent stuff if you knew how to pilot it)