Pros and cons of working on near-term technical AI safety and assurance — LessWrong