I hadn't heard the RLHF feedback loop argument before, but it seems like a solid explanation for all the bad ticks LLMs have when writing and why it's so hard to break them out of it.
All of that stuff... "That's not ____; it's _________" and the like... you can imagine human evaluators going "Wow, that's profound! Great output!" and now we all get to suffer through reading that for the rest of our lives.
All of that stuff... "That's not ____; it's _________" and the like... you can imagine human evaluators going "Wow, that's profound! Great output!" and now we all get to suffer through reading that for the rest of our lives.