An efficient, low-resolution machine learning model can usefully predict the global atmospheric state as much as three days out.