The benefits and risks of optimism (about AI safety) — LessWrong