Is there any plan to retry this with facebook's LLaMA models? They claim it outperforms GPT-3. Since it's open for research, it might give more precise answers to some of the questions.