Research conducted by Jerry Han at Princeton found that between 2018 and 2024, data improvements accounted for more than three times as much pre-training algorithmic progress as model architecture improvements