Alignment can improve generalisation through more robustly doing what a human wants - CoinRun example — LessWrong