I agree that we don't have any good statistical models for this. If AI development were that predictable we'd likely already be past a singularity of some sort or in a very long winter just by reverse-engineering what makes the statistical model tick.