If a language model is trained with 5x more compute than GPT-3 by the end of 2022, will it be public to end users?