Current AI scaling laws show that an exponential increase in compute is required to achieve a linear increase in performance, which argues against a sudden, fast takeoff