If it’s GENERIC reward seeking behaviour, why does alignment work sometimes? Why can we give a decent LLM a goal with a set of constraints and have it stay within those constraints fairly often?