Why do they not release the unnatural Variant of the model? According to the paper it beats all of the other variants and seems to be close to GPT-4.