How it works

Researchers trained models on synthetic data at multiple scales and measured how performance improved. They then compared those curves to runs on human-written data.