The GPT-4 technical report stated that OpenAI predicted GPT-4's final training loss using smaller models trained with 10,000 times less compute