Or, it’s a demonstration that AI output can be controlled in meaningful ways, period. Surely this supports openai’s stated goal of making safe AI?