I love/hate that most people do not understand how systems like LLMs work. Then they complain when it does things they want but to an extreme. Like someone said they wanted an AI to generate 59 videos, but instead it generated 500. Well, yes, that was 100% the possibility. The LLM said "My goal is to generate videos." so it did that as many times as it could. We've not created specific reward functions that prevent unwanted behaviour.