Designing agent incentives to avoid side effects — LessWrong