“The analogies for AI risk from Nick Bostrom and Eliezer Yudkowsky, such as the 'paperclipping' scenario, are outdated and not helpful for understanding the current alignment problem.”