Daniel Kokotajlo comments on Is GPT-3 already sample-efficient?