Until recently, training-time compute was the dominant cost, so we're really just getting started down the test-time scaling road.