Paradigm shifts in forecasting — LessWrong