gjm comments on PaLM in “Extrapolating GPT-N performance”