As always, these results are cherry picked by OpenAI. Same as GPT-2/3, the "average" output of the model is barely useable.