Limitations of reward clipping