Lanrian comments on Extrapolating GPT-N performance