The benefits, risks and bounds of personalizing the alignment of large language models to individuals