We're also increasingly using synthetic data to train them on, though, and the race now is in coming up with better ways to generate it.