“Future techniques for improving AI model quality will scale up inference compute, making a single request 50 to 1,000 times more computationally intensive.”