It's worth mentioning that they do RL for aesthetics to some degree based on human expert feedback, but whatever the model tends to produce quickly becomes debased by its ubiquity. They could RL for output diversity, but it's less well studied and likely to cause minor regressions in coding performance, at least until the algorithms are dialed in.