[aurareturn]

They have more data on what people want from models?

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