They are retrained every 12-24 months and constantly getting new/updated reinforcement learning layers. New concepts are not the problem. The problem is outdated information in the training data, like only crappy old Postgres syntax in most of the Stackoverflow body.