How often do the AIs devolve at some task? Or does switching models make those assumptions inaccurate?