gwern comments on Are Emergent Abilities of Large Language Models a Mirage? [linkpost]