RT @goodside: Every LLM dialog has an implicit prefix of “humans love reading shit like:” and the output is the model making fun of us.