Power laws in Speedrunning and Machine Learning — LessWrong