Bitter lesson is wrong. Large general models are outperformed by small specialized models.