I think most of the progress is training by reinforcement learning on automated assessments of the code produced. So data is not really an issue.