“The compute budget used for alignment fine-tuning of GPT-3 was less than 2% of its pre-training compute.”