“When Codex is working, users can send another message without waiting for the current run to finish to steer the model's output.”