“NVIDIA researchers have discovered a neural scaling law for dexterity, demonstrating a log-linear relationship between the hours of pre-training data and the model's optimal validation loss.”