Revisiting algorithmic progress — LessWrong