A framework for thinking about AI timescales — LessWrong