Governments should still be able to keep track of who's training frontier models — though a shift to reinforcement learning could make that harder