The commercial incentive to intentionally train AI to deceive us — LessWrong