“The seed data for training o1 was likely created over a year ago by high-skill human annotators who generated complex, rated reasoning paths.”