Logarithmic growth would probably be more accurate: every order of magnitude increase in model size tends to give a new ability