“It is not currently feasible to conduct ablation studies at the scale of GPT-4 to scientifically determine the impact of different data types on specific capabilities.”