“According to Epoch AI estimates, the training of GPT-4 used approximately 3,000 to 10,000 times more raw compute than the training of GPT-2.”