Imitative Generalisation (AKA 'Learning the Prior') — LessWrong