The GPT-4 technical report stated that final model loss was predictable using models trained with 10,000 times less compute