Hence the only other path: slow, constant enshittification of the models' performance