The alignment problem may be partially solved 'by default' through properties of large language models that naturally develop common sense understanding of goals from next-token prediction training, making them less prone to the Malign Omniscience failure modes that earlier alignment researchers feared.