Which shows that higher ups there don't understand how LLMs work. For one, negatives don't register well for them. And contradictory reponses just wash out as they work through repetition