PaLM-2 & GPT-4 in "Extrapolating GPT-N performance" — LessWrong