Their SOTA models can generate better synthetic data for the next training run - leading to a flywheel effect?
Their SOTA models can generate better synthetic data for the next training run - leading to a flywheel effect?