> The solution was going to be for everybody to write positive stories in which the LLM is good and relinquishes control, which then made it's way into the LLM's training data

Sounds like: https://www.lesswrong.com/w/aligned-ai-role-model-fiction