Forecasting future gains due to post-training enhancements — LessWrong